{
    "url": "http://arxiv.org/abs/2404.16385v1",
    "title": "Efficiency in Focus: LayerNorm as a Catalyst for Fine-tuning Medical Visual Language Pre-trained Models",
    "abstract": "In the realm of Medical Visual Language Models (Med-VLMs), the quest for\nuniversal efficient fine-tuning mechanisms remains paramount, especially given\nresearchers in interdisciplinary fields are often extremely short of training\nresources, yet largely unexplored. Given the unique challenges in the medical\ndomain, such as limited data scope and significant domain-specific\nrequirements, evaluating and adapting Parameter-Efficient Fine-Tuning (PEFT)\nmethods specifically for Med-VLMs is essential. Most of the current PEFT\nmethods on Med-VLMs have yet to be comprehensively investigated but mainly\nfocus on adding some components to the model's structure or input. However,\nfine-tuning intrinsic model components often yields better generality and\nconsistency, and its impact on the ultimate performance of Med-VLMs has been\nwidely overlooked and remains understudied. In this paper, we endeavour to\nexplore an alternative to traditional PEFT methods, especially the impact of\nfine-tuning LayerNorm layers, FFNs and Attention layers on the Med-VLMs. Our\ncomprehensive studies span both small-scale and large-scale Med-VLMs,\nevaluating their performance under various fine-tuning paradigms across tasks\nsuch as Medical Visual Question Answering and Medical Imaging Report\nGeneration. The findings reveal unique insights into the effects of intrinsic\nparameter fine-tuning methods on fine-tuning Med-VLMs to downstream tasks and\nexpose fine-tuning solely the LayerNorm layers not only surpasses the\nefficiency of traditional PEFT methods but also retains the model's accuracy\nand generalization capabilities across a spectrum of medical downstream tasks.\nThe experiments show LayerNorm fine-tuning's superior adaptability and\nscalability, particularly in the context of large-scale Med-VLMs.",
    "authors": "Jiawei Chen, Dingkang Yang, Yue Jiang, Mingcheng Li, Jinjie Wei, Xiaolu Hou, Lihua Zhang",
    "published": "2024-04-25",
    "updated": "2024-04-25",
    "primary_cat": "cs.CV",
    "cats": [
        "cs.CV"
    ],
    "label": "Original Paper",
    "paper_cat": "Parameter AND Efficient AND Fine AND Tuning",
    "gt": "Visual language models (VLMs) have become pivotal in facilitating multimodal tasks within the medical domain, such as medical visual question answering (Med-VQA) and medical imaging report gener- ation (Med-IRG). The pretraining-finetuning paradigm, heralded for its success in domain adaptation and transfer, now stands as the predominant training approach for Medical VLMs (Med-VLMs). Nonetheless, the substantial data and computational resource de- mands for VLMs\u2019 pretraining pose significant challenges. Despite the success of visual language pre-training paradigms like CLIP [32] and BLIP [21] fostering a series of open-source medical visual lan- guage pre-trained (VLP) models contributed by the community, adapting these models for specific downstream tasks remains a formidable task for those constrained by resource availability. Espe- cially considering the inherent minor variability in medical imaging across different medical centers or imaging devices, which necessi- tates researchers to frequently and rapidly fine-tune medical VLP models (Med-VLPs). The recent surge in Large Visual Language Models (LVLMs) has exacerbated these challenges. Although a series of Parameter- Efficient Fine-Tuning (PEFT) methods [13, 14, 25, 30] have been developed in the Large Language Model (LLM) domain, their ap- plicability and effectiveness in the context of LVLMs are yet to be ascertained[2]. Some empirical studies[36] have shown that the performance of certain PEFT methods contradicts their demon- strated competitiveness in the original LLM domain when fine- tuning domain-specific LVLMs because fine-tuning parameters for different parts of the LVLM can have drastically different effects. Distinct from universal domains, the medical field presents unique arXiv:2404.16385v1  [cs.CV]  25 Apr 2024 Preprint, 2024, Conference Paper Jiawei Chen, et al challenges, such as limited dataset size and specialized data char- acteristics, which necessitate a tailored approach to fine-tuning. These domain-specific requirements underscore the need for a ded- icated evaluation of PEFT methods on Med-VLMs to ensure their efficacy and appropriateness for medical tasks. Even so, research on the impact of imposing different PEFT methods on different scale Med-VLMs for different tasks remains completely blank. Mean- while, current PEFT methods typically focus on domain adaptation by adding extra components to the model structure (i.e., Reparame- terized Fine-tuning) [14, 50] or prefixes to the input (i.e., Additive Fine-tuning) [11, 25, 30], while the role of fine-tuning intrinsic struc- tural parameters of models has been widely neglected, especially in vertical domain fine-tuning. As shown in Figure 1, extrinsic tun- ing methods focus on concatenating additional components to the dense layer or adding prefixes to the inputs while intrinsic tuning methods select the intrinsic units of the transformer to be tuned. In this paper, we focus on efficiently fine-tuning Med-VLPs for specific downstream tasks, aiming to explore an innovative method that achieves task adaptation by fine-tuning a small subset of the model\u2019s intrinsic parameters. To find a universally efficient fine- tuning method applicable to various types of Med-VLMs, regardless of their scale, we turn our attention to common foundational lay- ers in transformer-based components, including Attention layers, Feed-Forward Neural Networks (FFN), and Layer Normalization (LayerNorm) layers. Through systematic experiments, we demon- strate that LayerNorm can serve as the most efficient method for fine-tuning Med-VLPs. To further explore the efficiency, adaptabil- ity, and performance of fine-tuning the intrinsic parameters for transferring Med-VLPs to downstream tasks, we conduct exten- sive evaluations across both large-scale and small-scale Med-VLPs, encompassing core medical visual language tasks like Med-VQA and Med-IRG. Our findings expose the variability of intrinsic pa- rameter fine-tuning methods in fine-tuning Med-VLMs to a down- stream task that is different from traditional extrinsic parameter fine-tuning methods. We underscore not only the superior efficiency of LayerNorm fine-tuning over existing PEFT methods but also its remarkable adaptability and transferability for fine-tuning Med- VLPs across diverse downstream tasks. The main contributions of this article are as follows: \u2022 To our best knowledge, we are the first to centre on fine- tuning a small subset of the Med-VLP\u2019s inherent parameters to adapt to downstream tasks. \u2022 We conduct a comprehensive series of experiments fine- tuning foundational components of Med-VLMs, including systematic comparisons with existing PEFT methods centred on tuning extrinsic components. \u2022 Our research identifies LayerNorm fine-tuning as a highly efficient strategy for adapting Med-VLPs to specific down- stream tasks within the medical domain.",
    "main_content": "In the medical domain, Med-VLMs play a pivotal role in automating visual-language tasks, such as VQA and IRG. Initially, these models [7, 10, 26, 31] leverage Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to extract visual and linguistic features separately. Yet, such approaches frequently fell short in terms of generalizability and transferability across different tasks due to the limitations of their structure. Driven by learning-based technologies [9, 17, 23, 24, 39\u201349], modern Med-VLMs [2\u20134] primarily adopt the transformer architecture, following a pretrainingfinetuning paradigm. They undergo initial pretraining on extensive, generalized medical image-text pair datasets, followed by comprehensive fine-tuning on more focused, task-specific datasets. For example, MISS [3], utilizing the ALBEF [22] methodology, begins its training on 38,800 selectively curated image-text pairs from the MedICaT dataset[33] before undergoing fine-tuning for VQA tasks. Similarly, LLaVA-Med [19] employs a dual-phase pretraining strategy, starting with image-text feature alignment on two million pairs from PubMed, then enhancing conversational capabilities using instruction-format data, culminating in full-scale fine-tuning for VQA tasks. These approaches consistently rely on full-model fine-tuning for task adaptation, a method that, despite its efficacy, demands substantial resources, particularly for large-scale models such as LLaVA-Med. The restricted dataset sizes available for downstream task training further jeopardize the model\u2019s generalizability, leading to potential catastrophic forgetting and diminishing its broader applicability in medical contexts. 2.2 Efficient Fine-tuning Techniques The fine-tuning of large-scale Pre-trained Language Models (PLMs) is a demanding process [1, 51, 53], requiring extensive computational resources and data. To alleviate these burdens, PEFT techniques [11, 25, 30, 36, 38] have been introduced. These methods [13, 14, 29] typically incorporate trainable components into the PLMs while maintaining the rest of the model\u2019s parameters in a frozen state. Some strategies [18, 25, 30] also involve the nuanced manipulation of input embeddings across different layers to minimize or negate modifications to the original model\u2019s architecture. PEFT methods have demonstrated efficacy in transitioning largescale PLMs to new tasks or downstream applications and have been instrumental in converting LLMs into multimodal LLMs [5, 20, 34, 51, 53]. For instance, LLaVA [28] uses an MLP adapter to connect a vision feature extractor with a large language model, selectively training the MLP adapter while keeping both components static, thus adapting the LLM into a VLM. [52] introduces an efficient strategy where tuning LayerNorm layers suffices to yield strong performance to transform an LLM into an LVLM. Nonetheless, the capability of existing PEFT methods to efficiently adapt pre-trained VLMs to specialized, especially medical, tasks remains largely uninvestigated. With the diverse architectures of LVLMs, the most effective application of PEFT methods is uncertain, and their generalizability to non-textual encoders/decoders is limited [12] (e.g., prefix-tuning and p-tuning are not viable for Vision Transformers (ViT) [8]). Consequently, investigating the adjustment of a model\u2019s intrinsic parameters for efficient fine-tuning emerges as a critical necessity. In this paper, we propose a novel method that eschews adding components to the original model structure or input, focusing instead on fine-tuning the model\u2019s inherent parameters. This strategy is designed to ensure the method\u2019s broad applicability for efficient fine-tuning across various Med-VLM types. Efficiency in Focus: LayerNorm as a Catalyst for Fine-tuning Medical Visual Language Pre-trained Models Preprint, 2024, Conference Paper 3 PRELIMINARIES 3.1 Mainstream Architectures of Med-VLMs Contemporary generative Med-VLMs, irrespective of their scale\u2014be it large-scale or small-scale, tend to follow a similar architectural framework. This typical structure comprises a vision feature extractor, a text feature extractor, a connector that integrates the former two, and a Language Model (LM) head. Most Med-VLMs opt for ViT [8] as the vision feature extractor, while the text encoder is based on mainstream frameworks such as BERT [6] or GPT [1]. Despite possible minor variations in their structural implementations, the transformer-based layer serves as their common denominator, with FFN, Attention mechanisms, and LayerNorm being indispensable core components. 3.2 Previous PEFT Methods Transitioning from the core mechanisms of attention and layer normalization, which provide stability and specificity within the model\u2019s architecture, we delve into the domain of extrinsic PEFT methods. These methods are categorized primarily into two types: Reparameterized Fine-tuning (i.e., LoRA (Low-Rank Adaptation)) and Additive Fine-tuning (i.e., Prefix-tuning). LoRA-Tuning: LLM maps data into a high-dimensional space for processing. LoRA indirectly trains the dense layers in the network by optimizing the rank-decomposition matrix that changes in the adaptation process of the dense layer, thereby achieving the best fine-tuning effect by optimizing only the rank-decomposition matrix of the dense layer. For the pretrained parameters \ud835\udf03\ud835\udc37 0 , the dense layer weight parameter matrix on a specific downstream task is defined as \ud835\udc4a0 \u2208R\ud835\udc51\u00d7\ud835\udc58and the intrinsic rank of it is \ud835\udf03\ud835\udc51; the specific downstream task\u2019s parameters \ud835\udf03\ud835\udc37is calculated as \ud835\udf03\ud835\udc37= \ud835\udf03\ud835\udc37 0 +\ud835\udf03\ud835\udc51\ud835\udc40, where \ud835\udc40is the rank-decomposition matrix. For\ud835\udc4a0 \u2208R\ud835\udc51\u00d7\ud835\udc58, LoRA updates it with the following equation: \ud835\udc4a0 + \u0394\ud835\udc4a= \ud835\udc4a0 + \ud835\udc35\ud835\udc34, \ud835\udc35\u2208R\ud835\udc51\u00d7\ud835\udc5f,\ud835\udc34\u2208R\ud835\udc5f\u00d7\ud835\udc58, (1) where \ud835\udc51is the output dimension of the previous layer, and \ud835\udc58is the input dimension of the next layer. For input \ud835\udc65, the forward propagation process is calculated as follows: \u210e= \ud835\udc4a0\ud835\udc65+ \u0394\ud835\udc4a\ud835\udc65= \ud835\udc4a0\ud835\udc65+ \ud835\udc35\ud835\udc34\ud835\udc65. (2) v v0Prefix-Tuning: Inspired by the In-Context Prompting method adopted by GPT3 [1], Li et al [25] propose the Prefix-tuning method for generation tasks. Instead of the discrete text used in prompt tuning, continuous vectors are prefixed to the input text. Specifically, the generation task is deemed as a table-to-text task, the input \ud835\udc65 is treated as a linear table and the output \ud835\udc66represents a short text. For an encoder-decoder model, different prefixes are attached to the beginning of the encoder and decoder with the input defined as: \ud835\udc67= [\ud835\udc43\ud835\udc45\ud835\udc38\ud835\udc39\ud835\udc3c\ud835\udc4b,\ud835\udc65, \ud835\udc43\ud835\udc45\ud835\udc38\ud835\udc39\ud835\udc3c\ud835\udc4b\u2032], and the prefixes are generated by a trainable matrix \ud835\udc43\ud835\udf03\u2208R|\ud835\udc43\ud835\udc56\ud835\udc51\ud835\udc65|\u00d7\ud835\udc51\ud835\udc56\ud835\udc5a(\u210e\ud835\udc56), the global training objective is defined as: max \ud835\udf19 log \ud835\udc43\ud835\udf19(\ud835\udc66|\ud835\udc65) = max \ud835\udf19 \u2211\ufe01 \ud835\udc56\u2208\ud835\udc4c\ud835\udc56\ud835\udc51\ud835\udc65 log \ud835\udc43\ud835\udf19(\ud835\udc67\ud835\udc56|\u210e<\ud835\udc56). (3) 3.3 Medical Visual Language Tasks Medical Visual Language Answering: The primary objective of Med-VQA is to provide answers based on professional questions posed by the inquirer regarding medical images, enhancing the understanding of medical images, and facilitating patient care through the automated interpretation of visual data. The tasks are categorized into open-ended questions, which require detailed descriptive answers, and close-ended questions, which demand concise, often binary responses like \u201cyes\u201d or \u201cno\u201d. This interdisciplinary domain requires the Med-VLM to interpret and provide insights into complex medical imagery, such as X-rays, MRI scans, and CT images. Medical Imaging Report Generation: Medical IRG involves the automatic creation of textual descriptions for medical images, using Med-VLMs. This task aims to analyze visual medical data and produce accurate, coherent, and clinically relevant reports. The goal is to assist radiologists and healthcare professionals by reducing their workload and improving diagnostic efficiency while maintaining high standards of patient care. 3.4 Difference between Universal VLMs and Med-VLMs\u2019 Downstream Tuning The differences between Med-VLMs and Universal VLMs make it necessary to study the tuning of Med-VLMs separately in addition to using the PEFT approach for Universal VLMs. From the data perspective, the dataset used for downstream task fine-tuning in the medical domain is extremely narrow compared to the universal domain, for example, the current largest radiological image dataset used for the Med-VQA task includes only 14,028 image-question pairs, which makes the fine-tuning of LVLMs fall into the problems of overfitting and catastrophic forgetting. At the same time, the answers of the textual Instruction pairs embedded in the dataset usually include only one or two simple words. Furthermore, the current training loss used by generative models in the fine-tuning process makes it very easy to fall into the learning of the data distribution from the long text to its short text rather than the learning of the correct image-text association. From the model perspective, most Med-VLMs are obtained by transfer learning from VLMs in the universal domain, a process where the visual coder is usually frozen, however, the domain gap between natural images and medical images affects the performance of Med-VLMs on medical tasks, where most of the current PEFT methods are not available for ViTs or or have not demonstrated their effective impact on visual encoders for transfer learning from general domain to medical domain. Therefore, a separate study of efficient fine-tuning methods for Med-VLM on downstream tasks is necessary. 4 TUNING SETTINGS 4.1 Baseline Model To explore a method that achieves task adaptation by fine-tuning a small subset of the model\u2019s own parameters, we choose two different-scale pre-trained Med-VLMs for different tasks: 1) smallscale VLM MISS [3] and 2) large-scale VLM LLaVA-Med [19] for Med-VQA and Med-IRG. These baseline models cover generative Med-VLM at different scales and for different tasks so that we can Preprint, 2024, Conference Paper Jiawei Chen, et al Assistant LayerNorm: The chest x-ray image shows a patient with a rightsided pneumothorax. A pneumothorax is a condition where air  accumulates in the pleural space, which is the area between the  lung and the chest wall.  This can cause the lung to collapse,  leading to difficulty breathing and chest pain. In this case, the  pneumothorax is located on the right side of the patient's chest. Attention: The chest x-ray shows no acute abnormalities. The  cardiac and mediastinal contours are normal, and there is no  evidence of pneumothorax or pleural effusion. The lungs are clear,  and there is no focal airspace consolidation. The impression is that  there is no acute cardiopulmonary abnormality. FFN: The chest x-ray shows no acute abnormalities. The heart size  and mediastinal contour are normal, and there is no evidence of  focal consolidation, pneumothorax, or pleural effusion.  Additionally,  there is no visible free intraperitoneal air under the diaphragm. The  impression is that there are no acute cardiopulmonary findings. LoRA: The chest x-ray image shows a patient with a right-sided  pleural effusion. A pleural effusion is an abnormal accumulation of  fluid in the pleural space, which is the area between the layers of  tissue that line the lungs and the chest cavity. The image displays a  silhouette of the heart and lungs, The presence of a pleural effusion  may cause the heart and lungs to appear differently than expected,  such as having a more pronounced silhouette. Prefix: The chest x-ray image shows a patient with a right-sided  pneumothorax. A pneumothorax is a condition where air  accumulates in the pleural space, which is the area between the  lung and the chest wall. This can cause the lung to collapse,  leading to difficulty breathing and chest pain. Question: Describe the given chest x-ray image in detail. Question: What modality is used to take this image? (Open-ended) Assistant LayerNorm: The image is taken using a computed tomography (CT)  scan. Attention: Computed tomography (CT) is the modality used to take  this image. FFN: CT scan, or computed tomography scan, is the modality used to  take this image. LoRA: The image is taken using a computed tomography (CT) scan. Prefix: The image is taken using a computed tomography (CT) scan. Question: Does this image look normal? (Closed-ended) Assistant LayerNorm: No, the image does not appear to be normal. It shows  bilateral infiltrates, which are abnormal substances or collections of  cells within the lung tissue. This can be indicative of various  conditions, such as infection, inflammation, or other lung diseases. Attention: No, this image does not look normal. It shows bilateral  patchy infiltrates, which are abnormal areas of increased opacity in  the lungs. FFN: No. LoRA: No, this chest X-ray does not look normal. It shows bilateral  infiltrates, which are abnormal opacities or densities in both lungs. Prefix: No, the image does not appear to be normal. It shows  bilateral infiltrates, which are abnormal substances or collections of  cells within the lung tissue. This can be indicative of various  conditions, such as infection, inflammation, or other lung diseases. Vision Encoder Image Embeddings Language model Question \u201cIs there any  pneumonia in  the X-ray?\u201d Ground-truth \u201cYes\u201d Connector LM head Text Output Prompt for Eval You will act as an intelligent evaluator of answers generated by Generative Medical  Visual Language Model (Med-VLM). Please note that Med-VLM answers may be  more varied than benchmark answers. If a Med-VLM answer is approximately correct  about the image from a medical point of view, it should be regarded as correct even  if there are some differences from the benchmark answer, so do not arbitrarily give  an incorrect assessment. I'll give you the $questions$, $Med-VLM answer$, and  $ground_truth$. You must output a word $correct$ or $incorrect$: $questions$: What modality is used to take this image?  $Med-VLM answer$: The image is taken using a computed tomography (CT) scan. $ground_truth$: CT. or Figure 2: The pipeline of our study. The flowchart details the step-by-step process from input reception to output generation, showcasing the model\u2019s method for processing medical images and questions to generate contextually relevant responses. The right side presents the fine-tuning results across different paradigms, including both Med-VQA and Med-IRG tasks. provide comprehensive insights into the impact of different PEFT methods on fine-tuning Med-VLM to downstream tasks. 4.2 Tuning within Transformer-based Layer The transformer-based layer, serving as a fundamental structure across vision encoders, language models, and certain VLM connectors, undergoes fine-tuning through both intrinsic parameter adjustments and the application of extrinsic component fine-tuning methods. This process aims to assess the impact of various tuning approaches on the overall model performance. Strategic Intrinsic Adjustments: We emphasize the transformerbased layer\u2019s role as the computational core of the model and selectively fine-tune its intrinsic parameters. Attention layers, LayerNorm layers, and FFNs constitute the critical units of this layer. As depicted in Figure 1, when one of these three components is set to be trainable, the remaining parameters within the transformer-based layer are kept frozen. Incorporation of PEFT Techniques: LoRA-tuning and Prefixtuning, representing the forefront of PEFT methods, are chosen for comparison against intrinsic parameter adjustments. Figure 1 illustrates the application details of these PEFT methods within our study. For LoRA-tuning, low-rank matrices are selectively applied to the parameters of the query and value matrices within the attention layer, mapping data into a low-dimensional subspace for processing. In the case of Prefix-tuning, we follow prevalent practices by appending prefix vectors to the origin input \ud835\udc65of the key and value within the attention layer, and the final input embeddings can be defined as \ud835\udc67= [\ud835\udc43\ud835\udc45\ud835\udc38\ud835\udc39\ud835\udc3c\ud835\udc4b,\ud835\udc65]. 4.3 Tuning within the VLM Architecture In our pursuit to uncover the nuanced impact of various modules within the full VLM\u2019s architecture on overall model performance, we embark on a strategic fine-tuning expedition which entails selectively training specific modules within the VLM framework while employing efficient fine-tuning methods for certain components or maintaining others in a frozen trainable state. Such a strategy allows us to dissect the individual contributions of each component to the model\u2019s efficacy in medical visual language tasks, offering insights into optimizing Med-VLMs for enhanced performance and efficiency. This selective fine-tuning approach aims to validate the hypothesis that certain components within the Med-VLM architecture wield more significant influence over the model\u2019s performance on medical tasks. By applying focused fine-tuning strategies to individual modules, we seek to delineate the performance impact of targeted adjustments versus broad model updates. Details for Small-scale Med-VLMs: For small-scale Med-VLMs, such as MISS, we experiment with applying efficient fine-tuning techniques to some modules at a time, with the rest of the model\u2019s parameters set to remain fully trainable This is because for smallscale Med-VLM, making either module completely frozen may make the model unable to transfer to downstream tasks [2]. For instance, when the language model undergoes LayerNorm-tuning (LN-tuning), the vision feature extractor, connector, and LM head are kept in a state that allows full parameter adjustments. This strategy allows us to compare the impact of different efficient finetuning methods on model performance under the premise that we can evaluate the impact of fine-tuning different module parameters of small-scale Med-VLMs on the overall task performance. Efficiency in Focus: LayerNorm as a Catalyst for Fine-tuning Medical Visual Language Pre-trained Models Preprint, 2024, Conference Paper Strategy for Large-scale Med-VLMs: In the case of large-scale Med-VLMs, like LLaVA-Med, our fine-tuning strategy is more nuanced, reflecting the diverse requirements of comparative analysis. As shown in Figure 2, \u201cSnowflakes or flames\u201d indicate that any module has the option to be adjusted or frozen. Depending on the specific experimental setup, modules within these larger models may be categorized into three states: fully trainable (T), efficiently fine-tuned some of the parameters (PEFT), and completely frozen (F). This flexible approach comprehensively evaluates how different tuning states across various modules influence large-scale VLMs\u2019 performance on complex medical visual language tasks. 4.4 Downstream Fine-tuning Benchmarks: In this paper, we use a total of four datasets, Slake Dataset [27] and VQA-RAD dataset [16] for Med-VQA, and OpenI dataset [37] and MIMIC dataset[15] for Med-IRG. The Slake dataset consists of 14,028 QA pairs, of which 70% are used for training, 15% for validation, and 15% for testing. The VQA-RAD dataset is used for the zero-shot performance of the model on the VQA task including 3515 QA pairs, of which 451 pairs are used for testing. The OpenI dataset is used for the training of the Med-IRG task including 6,459 images and 3,955 reports, the instructions are the same as those adopted in [35]. The MIMIC test set was chosen for the evaluation of the model\u2019s Med-IRG performance, which includes 5,159 images and 3,269 reports. For the inference of the IRG task, we uniformly use the phrase \u201cDescribe the given chest x-ray image in detail.\u201d as the instruction. Application Details: For the fine-tuning of our chosen models (MISS and LLaVAMed), a consistent set of hyperparameters is employed to ensure uniformity across our experiments. Each model is fine-tuned with an initial learning rate of 2e-5, utilizing the Adam optimizer for its well-regarded efficiency in handling the optimization landscapes of deep learning models. Specifically, MISS underwent training for 120 epochs with a batch size of 16, adopting a weight decay of 0.05 to encourage regularization. In contrast, LLaVA-Med\u2019s finetuning is characterized by a warmup ratio of 0.03 and a cosine learning rate scheduler, alongside specific adjustments such as enabling tensor float 32 for enhanced computational performance, and employing FSDP strategies for memory efficiency, with settings like \u201cfull_shard auto_wra\u201d and targeting the \u201cLlamaDecoderLayer\u201d for wrapping. During all the inferences, Med-LLaVA generates outputs using a set of predefined generation parameters, including sampling methods and beam search configurations (num-beams=1), and the temperature is kept at 0.2. Detailed information regarding the hyperparameter settings for each model, along with additional configurations and the rationale for their selection, is provided in the Appendix for further reference. 5 EXPERIMENT RESULTS AND DISCUSSION 5.1 Small-scale MISS Result As shown in Table 1, we employ both supervised fine-tuning (SFT) and performance assessment on the Slake dataset\u2019s training and testing sets. The reported performance metrics include accuracy rates for \u2018opened\u2019 and \u2018closed\u2019 types, which means open-ended and closed-ended questions, as well as a global accuracy rate that averages the performance across both types. Given the potentially catastrophic impact of freezing any module on the overall performance of small-scale models, when certain modules underwent efficient fine-tuning, the remaining modules were maintained fully trainable. In the context of the MISS model, ViT, JTM, and DEC represent the visual encoder, joint textmultimodal encoder, and text decoder, respectively. The term \u201ctrainable params\u201d refers to the total volume of trainable model parameters, with \u201c#Params\u201d indicating the ratio of trainable to total parameters. \u201cPEFT params\u201d denotes the proportion of parameters fine-tuned using PEFT methods, with \u201c#PEFT Params\u201d reflecting the proportion of PEFT-tuned parameters relative to the total parameters within the corresponding module. When the baseline model is fully fine-tuned, it achieves the highest open-ended question accuracy and global accuracy rates of 82.91% and 82%, respectively. Under the premise of maintaining the visual encoder fully trainable and only efficiently tuning one module at a time, keeping the JTM encoder fully trainable enabled the model to achieve optimal performance. Compared to scenarios where DEC underwent full parameter training while JTM was efficiently tuned, the model\u2019s global accuracy rates under LayerNorm, attention, and FFN intrinsic parameter tuning methods were higher by 4%, 9%, and 12%, respectively. Maintaining the visual encoder fully trainable while efficiently tuning all the remaining modules resulted in significantly poor model performance, failing to correctly judge the closed-source questions in all the fine-tuning methods except Attention-tuning. Under LayerNorm, FFN, LoRA, and Prefix-tuning methods, the model never answered \u2018yes\u2019 to any close-ended question, with accuracy rates lower than random guessing at 38.03%, 54.65%, 50.70%, and 32.95%, respectively. Comparing different fine-tuning methods, the effect of LN-tuning is remarkable, achieving the best accuracy on close-ended questions with the lowest PEFT Params, even surpassing full parameter tuning and reaching 84.51%. In contrast, although Attention-tuning and FFN-tuning slightly outperform LN-tuning in terms of global accuracy, this came at the cost of tuning over 40% of the parameters in their respective modules. The LoRA method fine-tuning model using the [T, PEFT, T] paradigm tuned approximately five times more PEFT parameters than LN-tuning (only 56,823), with Prefix-tuning at twenty times more. This underscores the viability of LN-tuning as a comparable method to the most classical PEFT methods in small-scale fine-tuning scenarios that require saving certain parameter volumes. From a global parameter tuning perspective, Attention-tuning achieves performance closest to full fine-tuning by saving 23% of trainable parameters, marking it as another viable fine-tuning approach for small-scale Med-VLMs. 5.2 Large-scale LLaVA-Med Result Furthermore, we conduct comprehensive evaluations on LLaVAMed, a large-scale model designated for Med-VQA tasks. Our approach encompassed four distinct training paradigms: [PEFT, F, PEFT, F], [PEFT, T, PEFT, T], [F, F, PEFT, F], and [F, T, PEFT, T]. Considering the substantial parameter size of LVLMs, we aimed to restrict the volume of fine-tuning parameters to within about Preprint, 2024, Conference Paper Jiawei Chen, et al Table 1: Comparison of accuracy (ACC-%) of MISS on Slake dataset using different methods of fine-tuning. \u2018T\u2019 stands for trainable while \u2018F\u2019 stands for frozen. ViT JTM DEC Opened \u2191 Closed \u2191 Gobal \u2191 Trainable Params #Params PEFT Params #PEFT Params T T T 82.91 81.47 82.00 361,478,972 100% T LayerNorm LayerNorm 40.79 38.03 39.87 86,454,528 23.92% 115,200 0.04% T LayerNorm T 75.64 84.51 78.61 224,277,308 62.04% 56,832 0.04% T T LayerNorm 73.65 77.46 74.93 223,656,192 61.87% 58,368 0.04% T Attention Attention 64.51 74.65 71.25 199,806,720 55.27% 113,467,392 41.24% T Attention T 78.47 85.92 80.96 280,954,172 77.72% 56,733,696 41.33% T T Attention 75.50 64.23 71.72 280,331,520 77.55% 56,733,696 41.15% T FFN FFN 74.79 54.65 68.05 199,677,696 55.24% 113,338,368 41.19% T FFN T 76.63 84.51 79.26 280,889,660 77.71% 56669184 41.27% T T FFN 76.20 49.86 67.39 280,267,008 77.53% 56669184 41.10% T LoRA LoRA 68.14 50.70 62.29 86,929,152 24.05% 589,824 0.21% T LoRA T 76.77 82.81 78.79 224,515,388 62.11% 294,912 0.21% T T LoRA 78.52 79.44 78.83 223,892,736 61.94% 294,912 0.21% T Prefix Prefix 41.50 32.95 38.61 115,884,288 32.06% 29,544,960 10.74% T Prefix T 75.92 83.38 78.42 238,992,956 66.12% 14,772,480 10.76% T T Prefix 76.82 82.25 78.65 238,370,304 65.94% 14,772,480 10.71% Table 2: Comparison of results of LLaVA-Med on the Slake dataset using different methods of fine-tuning. Vision Tower Connector LLM LM Head Slake Dataset #Params Trainable Params BERTS-Recall Mean Token Opened \u2191 Closed \u2191 Global \u2191 LayerNorm F LayerNorm F 59.53 69.95 63.62 3.79% 266,737,664 46.35% 28.27 F F LayerNorm F 58.76 69.71 63.05 0.00372% 262,144 46.24% 27.81 LayerNorm T LayerNorm T 59.84 67.55 62.87 3.79% 266,737,664 46.93% 26.81 F T LayerNorm T 60.31 66.11 62.58 3.78% 266,637,312 46.93% 26.50 Attention F Attention F 61.4 67.79 63.9 31.91% 2,248,245,248 49.25% 25.24 F F Attention F 61.71 68.03 64.18 30.48% 2,147,483,648 49.11% 25.95 Attention T Attention T 60.93 65.87 62.87 35.69% 2,514,620,416 48.47% 25.89 F T Attention T 58.76 66.83 61.92 34.26% 2,413,858,816 48.49% 25.85 F T FFN T 64.5 62.26 63.62 44.74% 3,152,056,320 51.98% 16.42 F F FFN F 64.34 66.59 65.22 40.96% 2,885,943,296 52.07% 17.37 F F LoRA F 58.14 64.42 60.6 0.14% 9,994,240 47.55% 25.43 F T LoRA T 58.76 65.38 61.36 3.92% 276,369,408 47.26% 25.72 F F Prefix F 56.9 67.07 60.89 15.48% 1,090,805,760 46.19% 26.61 F T Prefix T 59.22 70.19 63.52 19.26% 1,357,180,928 46.28% 26.58 40%, thereby excluding full parameter training of the ViT and FFNtuning methods that involve adjusting ViT. Table 2 showcases the experimental results of LLaVA-Med, trained and tested on the Slake dataset, employing the aforementioned fine-tuning paradigms. When opting to keep both the connector and LM head trainable, the model\u2019s performance did not exhibit significant improvement, despite a substantial increase in the volume of adjusted parameters. Specifically, when fine-tuning adopted the [F, T, PEFT, T] paradigm, changes in global accuracy rates for LN, attention, and FFN tuning compared to [F, F, PEFT, F] are -0.5%, -2.1%, and -3.3% respectively. This contradicts the common notion that more parameter adjustments correlate with better SFT performance, indicating that full parameter adjustments of the connector and LM head during efficient fine-tuning of LLMs do not guarantee the expected outcomes. The performance changes are inconsistent under the [PEFT, F, PEFT, F] and [PEFT, T, PEFT, T]. For LN-tuning, fine-tuning the image encoder led to respective increases in global accuracy of 0.57% and 0.29%, while Attention-tuning resulted in changes of -0.28% and +0.85%. Such subtle differences do not conclusively indicate whether adjusting parameters of the image encoder benefits or hinder model performance, especially when considering Recall metrics. The increase in ViT-adjusted parameter volume did not regularly alter recall, suggesting that larger adjustments to ViT parameters do not consistently improve model recall. Comparing different intrinsic parameter adjustments revealed that increasing the volume of fine-tuned parameters indeed enhances the model\u2019s recall of generated content: as fine-tuning parameters shifted from 0.003% to 44%, recall correspondingly increased from 46.24% to 52.07%. This indicates that enlarging the volume of fine-tuned parameters allows the model to learn the distribution of ground-truth tokens in the vocabulary space more effectively, both quantitatively and spatially. However, considering accuracy\u2014a gold standard in medical tasks\u2014significant increases in parameter volume do not necessarily elevate all accuracy metrics concurrently. LN-tuning under the [PEFT, F, PEFT, F] paradigm once again achieved state-of-the-art (SOTA) accuracy for close-ended questions, which was consistent with observations in small-scale VLMs. Across two models of different scales, pre-trained on distinct Efficiency in Focus: LayerNorm as a Catalyst for Fine-tuning Medical Visual Language Pre-trained Models Preprint, 2024, Conference Paper Figure 3: Loss curves for various methods under the [F, T, PEFT, T] fine-tuning paradigm. datasets and tasks, LN-based fine-tuning consistently enhanced their accuracy on close-ended questions. While Attention-tuning and FFN-tuning marginally surpassed LN-tuning in global accuracy, achieving peak open-ended question accuracies of 64.34% and global accuracy of 65.22%, this came at the cost of escalating the volume of tuned parameters from 262,144 to 2,885,943,296\u2014a millionfold increase. Furthermore, following peak performance under current fine-tuning paradigms, the model ceased learning intrinsic relations of the features, instead focusing on the quantitative distribution of ground-truth tokens. This shift manifested in minimal accuracy improvements and a dramatic reduction in average output length, with mean tokens dropping from 28.27 to 17.38. Figure 2 compares the generated outcomes across different intrinsic and extrinsic fine-tuning methods under the [F, F, PEFT, F] paradigms, illustrating this phenomenon driven by inherent large model training patterns and suboptimal training data. Most datasets employed for fine-tuning Med-VQA tasks comprise answers in short text formats, with close-ended answers typically being \u2018yes\u2019 or \u2018no\u2019, and open-ended answers containing only a few words. For LLMs seeking interpretability of the answer, adjusting more parameters paradoxically impairs generative performance. When applying LoRA-tuning and Prefix-tuning to LLaVA-Med, the model\u2019s performance did not exhibit notable improvement. LoRA-tuning\u2019s recall reached 47.55% and 47.26%, indicating a closer alignment of the model\u2019s output with the ground truth distribution in the vocabulary space. However, this did not translate to enhanced evaluation accuracy, with accuracies under [F, F, PEFT, F] and [F, T, PEFT, T] fine-tuning paradigms reaching only 60.6% and 61.36%, respectively. This suggests that LoRA-tuning failed to deepen the multimodal model\u2019s understanding of joint image-text features, merely aligning output closer to the ground truth distribution. Figure 3 displays the loss curves for various methods under the [F, T, PEFT, T] fine-tuning paradigm. Compared to LN-tuning, LoRAtuning\u2019s minimum fine-tuned parameter volume is approximately fifty times larger, yet its accuracy is roughly 3% lower, with average output lengths of 25.43 and 25.72, trailing behind LN-tuning. These factors collectively indicate that existing PEFT methods may not directly enhance text-based accuracy in discerning medical images in multimodal model downstream fine-tuning, underscoring the advantages of LN-tuning over traditional PEFT approaches. Table 3: Comparison of LLaVA-Med competence on the MIMIC test dataset which SFT on the OpenI dataset. Vision Tower Connector LLM LM Head METEOR Score Rouge-L Mean Token Recall Precision F1 F F LayerNorm F 12.85% 12.58% 15.92% 13.61% 122.66 F F Attention F 12.85% 12.58% 15.92% 13.61% 122.66 F F FFN F 24.53% 17.01% 23.84% 19.34% 123.11 F F LoRA F 12.95% 12.57% 15.93% 13.62% 122.70 F F Prefix F 12.99% 12.47% 15.92% 13.54% 122.72 5.3 Large-scale VLM IRG Result To further explore the impact of various fine-tuning methods on the performance of large-scale Med-VLMs in the Med-IRG context, we employed the [F, F, PEFT, F] fine-tuning paradigm using the OpenI dataset to fine-tune Med-LLaVA. The model\u2019s OOD performance was then tested on the MIMIC dataset\u2019s test set to assess how it handles variations in input text domains. As shown in Table 3, the performance of the models fine-tuned with LN, Attention, LoRA, and Prefix methods showed minimal differences, with output text lengths averaging around 122.7. In contrast, FFN tuning significantly outperformed other fine-tuning approaches, demonstrating its superior capability in learning the underlying representations of ground-truth in long text generation tasks like Med-IRG. 5.4 Out of Distribution Performance Testing To assess whether the performance of LLaVA-Med on a familiar dataset like Slake correlates with its performance on a novel dataset, we conducted an OOD testing on the VQA-RAD dataset. This test serves to evaluate the model\u2019s robustness and flexibility by applying it to a different domain within the same field but with unseen data. More specifically, the images in the VQA-RAD dataset belong to the proximity domain with the Slake dataset but are quite different from the Slake dataset in terms of question formulation. Such experiments allow us to consider the ability of different fine-tuning methods to reason on non-proximity-domain text over similar medical images, in order to speculate on the ability of the models fine-tuned with the VQA dataset to be applied to the real Med-VQA scenarios. In this experiment, we observe various fine-tuning paradigms, focusing particularly on the role of the transformer-based LayerNorm, Attention, and FFN adjustments. The results show a notable variance in the model\u2019s ability to generalize the learned features to the VQA-RAD dataset. As Table 4 shows, fine-tuning methods that show comparable results on the Slake dataset exhibit significant performance variances on the OOD VQA-RAD dataset. Notably, under the [F, T, PEFT, T] training paradigm, LoRA-tuning underwent a remarkable reversal, surpassing the performances of Attentiontuning and FFN-tuning, which were previously effective on the Slake dataset. It achieves a global accuracy of 65.41% and matches the best-closed question accuracy of 73.71%, initially noted with LNtuning. Conversely, FFN-tuning, despite being superior at learning adjacent training text representations, disappointed in its OOD performance. While it excels in Rouge-L metrics with scores of 26.35%, 9.80%, and 13.51%, significantly surpassing other fine-tuning methods, it only managed accuracy scores of 56.50%, 64.54%, and 60.98%. Moreover, its mean output length plummeted to 18.51, the lowest among all methods. This combination of metrics further validates that although FFN-tuning can closely fit the training data during Preprint, 2024, Conference Paper Jiawei Chen, et al Table 4: Comparison of LLaVA-Med competence on the VQA-RAD test dataset which SFT on the Slake dataset. Vision Tower Connector LLM LM Head VQA-RAD Bertscore METEOR Score Rouge-L Mean Token Opened\u2191 Closed\u2191 Global\u2191 Precision Recall F1 Recall Precision F1 F T LayerNorm T 54.50 73.71 65.19 29.77% 49.00% 36.53% 12.53% 7.91% 1.48% 2.30% 29.52 F T Attention T 50.50 63.35 57.65 29.97% 49.27% 36.73% 11.87% 7.97% 1.46% 2.30% 31.12 F T FFN T 56.50 64.54 60.98 35.44% 52.05% 41.89% 19.23% 26.35% 9.80% 13.51% 18.50 F T LoRA T 55.00 73.71 65.41 30.06% 49.52% 36.87% 12.03% 7.89% 1.50% 2.29% 27.52 F T Prefix T 51.00 70.12 61.64 29.89% 48.85% 36.58% 12.71% 8.53% 1.57% 2.45% 28.87 F F LayerNorm F 54.50 75.30 66.08 29.90% 48.84% 36.60% 12.97% 7.84% 1.50% 2.33% 29.52 F F Attention F 55.50 71.71 64.52 30.25% 49.64% 37.08% 8.44% 8.44% 1.66% 2.58% 29.45 F F FFN F 52.00 61.35 57.21 35.20% 51.76% 41.60% 18.43% 24.31% 8.96% 12.40% 19.16 F F LoRA F 49.50 70.92 61.42 30.08% 49.80% 36.96% 12.46% 7.91% 1.51% 2.33% 29.04 F F Prefix F 51.00 69.72 61.42 29.76% 48.49% 36.36% 12.81% 8.11% 1.54% 2.41% 29.21 Table 5: Comparison of LLaVA-Med zero-shot competence on the MIMIC test dataset which SFT on the Slake dataset. Vision Tower Connector LLM LM Head METEOR Score Rouge-L Mean Token Recall Precision F1 F T LayerNorm T 11.86% 11.27% 18.54% 13.54% 71.78 F T Attention T 11.88% 11.15% 17.98% 13.33% 73.73 F T FFN T 12.12% 11.24% 18.07% 13.42% 73.23 F T LoRA T 11.57% 10.99% 17.91% 13.18% 71.51 F F LayerNorm F 12.07% 11.40% 18.26% 13.56% 75.00 F F Attention F 12.17% 11.50% 18.20% 13.65% 75.01 F F FFN F 12.81% 11.48% 17.84% 13.53% 76.88 F F LoRA F 11.91% 11.25% 18.07% 13.40% 72.93 SFT, it predominantly learns the distribution of ground-truth tokens rather than enhancing the model\u2019s ability to generalize image-text reasoning. A similar pattern is observed with Attention-tuning; the global accuracy decreases by 7.54% compared to LN-tuning when the connector and LM head are trained more. However, the text length is optimal at this time. Comprehensively, the mean token length of attention tuning under the same paradigm in Table 1 can show that attention tuning slightly overlearns the text in the adjacent domains during SFT training, but does not cause large damage to the model\u2019s output ability of textual diversity in textual reasoning in the non-adjacent domains. When enlarging the perspective to compare the accuracy performances across different fine-tuning methods, LN-tuning consistently displayed formidable strength. Under the [F, F, PEFT, F] tuning paradigm, LN-tuning, utilizing the smallest parameter adjustment, reached the highest scores in opened, closed, and global accuracy at 54.5%, 75.3%, and 66.08%, respectively. It also maintains a longer mean text output than any other method under the same tuning conditions. These results, coupled with the Rouge-L metrics from Table 1, indicate that LN-tuning manages to enhance the model\u2019s understanding of multimodal feature interrelations, significantly minimizing the model\u2019s overemphasis on learning groundtruth text token distributions due to low training data quality. This is evidenced by the lowest recall rate of 1.5% and the highest global accuracy of 66.08%. In contrast, LoRA-tuning, despite adjusting 50 times more parameters, did not significantly outperform LN-tuning. 5.5 Zero-shot Capability Investigation To further investigate the zero-shot capabilities of different finetuning paradigms on Med-VLMs, we conduct extensive evaluations of the LLaVA-Med model on the MIMIC test dataset after SFT on the Slake dataset. This analysis aims to understand the impact of various intrinsic tuning methods on the model\u2019s ability to generalize and adapt to new tasks within the medical domain, particularly for IRG tasks. The evaluation employs metrics such as METEOR score, Rouge-L, and mean token length to measure factual accuracy, linguistic precision, and diversity of output in medical report generation. From Table 5, experimental results indicate that different tuning methods exhibit varying impacts on the model\u2019s zero-shot performance. LN-tuning consistently showed robust performance across different configurations, achieving the highest precision scores (18.54% and 18.26% under different paradigms), which underscores its effectiveness in preserving the factualness of model outputs. In contrast, the Attention and FFN methods, although effective in some scenarios, demonstrate greater variability in their influence on model generalization. Notably, FFN-tuning, which previously excelled in VQA tasks, scored the lowest in precision (17.84%) under the [F, F, FFN, F] paradigm on the MIMIC dataset. This suggests that the model may have overlearned task-specific features from the Slake dataset, thus hindering its generalization and transfer capabilities. Furthermore, METEOR scores positively correlated with the number of adjusted parameters, increasing from 11.86% to a high of 12.81%, indicating that a larger volume of tuned parameters enhances the model\u2019s linguistic alignment capabilities in medical text generation tasks. Examining the effects of freezing versus tuning the connector and LM head reveals no clear pattern in performance metrics between [F, T, PEFT, T] and [F, F, PEFT, F] configurations. Changes in Rouge-L score and precision are minimal, suggesting that extensive finetuning of the connector and LM head does not necessarily contribute to improved zero-shot performance across these metrics. Comparison between intrinsic tuning and traditional methods, such as LoRA-tuning, did not exhibit standout performance in the zero-shot setting. Under the [F, T, PEFT, T] paradigm, LoRA-tuning shows lower METEOR scores, Rouge-L, and mean token length compared to intrinsic methods, indicating that LoRA-tuning might not effectively maintain the overall transferability and generalization of the model in medical applications. Thus, intrinsic tuning methods, particularly LN-tuning with minimal parameter adjustments, might be a better choice, especially under the [F, F, PEFT, F] paradigm, where it outperforms more parameter-intensive methods like Attention-tuning in maintaining the generalization capabilities. These observations underscore the efficacy of LN-tuning in preserving the generalization of LVLMs for diverse medical tasks. Efficiency in Focus: LayerNorm as a Catalyst for Fine-tuning Medical Visual Language Pre-trained Models Preprint, 2024, Conference Paper 6 CONCLUSION This study presents a thorough examination of intrinsic parameter fine-tuning and exposing LN-tuning, as a potent alternative to traditional PEFT methods for Med-VLMs. Our extensive experimental analysis across both small-scale and large-scale Med-VLMs demonstrated that fine-tuning the LayerNorm layers significantly enhances the models\u2019 adaptability, efficiency, and scalability in performing specialized medical tasks, such as Med-VQA and MedIRG. We hope this work will enhance the clinical applicability of Med-VLMs in real-world medical settings. ACKNOWLEDGMENTS This work is supported in part by the National Key R&D Program of China (2021ZD0113503) and in part by the Shanghai Municipal Science and Technology Committee of Shanghai Outstanding Academic Leaders Plan (No. 21XD1430300). REFERENCES [1] Tom Brown, Benjamin Mann, Ryder, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877\u20131901. [2] Jiawei Chen, Yue Jiang, Dingkang Yang, Mingcheng Li, Jinjie Wei, Ziyun Qian, and Lihua Zhang. 2024. Can LLMs\u2019 Tuning Methods Work in Medical Multimodal Domain? arXiv preprint arXiv:2403.06407 (2024). [3] Jiawei Chen, Dingkang Yang, Yue Jiang, et al. 2024. MISS: A Generative Pretraining and Finetuning Approach for Med-VQA. arXiv preprint arXiv:2401.05163 (2024). [4] Zhihong Chen, Yuhao Du, Hu, et al. 2022. Multi-modal masked autoencoders for medical vision-and-language pre-training. In MICCAI. Springer, 679\u2013689. [5] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. [n. d.]. InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv 2023. arXiv preprint arXiv:2305.06500 ([n. d.]). [6] Jacob Devlin, Ming-Wei Chang, Lee, et al. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018). [7] Tuong Do, Binh X. Nguyen, et al. 2021. Multiple Meta-model Quantifying for Medical Visual Question Answering. In MICCAI. Cham, 64\u201374. [8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020). [9] Yangtao Du, Dingkang Yang, Peng Zhai, Mingchen Li, and Lihua Zhang. 2021. Learning Associative Representation for Facial Expression Recognition. In IEEE International Conference on Image Processing (ICIP). 889\u2013893. [10] Haifan Gong, Guanqi Chen, Mao, et al. 2022. VQAMix: Conditional Triplet Mixup for Medical Visual Question Answering. IEEE Transactions on Medical Imaging 41, 11 (2022), 3332\u20133343. https://doi.org/10.1109/TMI.2022.3185008 [11] Zeyu Han, Chao Gao, Jinyang Liu, Sai Qian Zhang, et al. 2024. ParameterEfficient Fine-Tuning for Large Models: A Comprehensive Survey. arXiv preprint arXiv:2403.14608 (2024). [12] Xuehai He, Chunyuan Li, Pengchuan Zhang, Jianwei Yang, and Xin Eric Wang. 2023. Parameter-efficient model adaptation for vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 817\u2013825. [13] Neil Houlsby, Andrei Giurgiu, Jastrzebski, et al. 2019. Parameter-efficient transfer learning for NLP. In International Conference on Machine Learning. PMLR, 2790\u2013 2799. [14] Edward J Hu, Shen, et al. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021). [15] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. 2019. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data 6, 1 (2019), 317. [16] Jason J Lau, Gayen, et al. 2018. A dataset of clinically generated visual questions and answers about radiology images. Scientific data 5, 1 (2018), 1\u201310. [17] Yuxuan Lei, Dingkang Yang, Mingcheng Li, Shunli Wang, Jiawei Chen, and Lihua Zhang. 2023. Text-oriented Modality Reinforcement Network for Multimodal Sentiment Analysis from Unaligned Multimodal Sequences. arXiv preprint arXiv:2307.13205 (2023). [18] Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021). [19] Chunyuan Li, Cliff Wong, Zhang, et al. 2023. Llava-med: Training a large languageand-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890 (2023). [20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023). [21] Junnan Li, Dongxu Li, Xiong, et al. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICCV. 12888\u201312900. [22] Junnan Li, Ramprasaath Selvaraju, Gotmare, et al. 2021. Align before fuse: Vision and language representation learning with momentum distillation. NIPS 34 (2021), 9694\u20139705. [23] Mingcheng Li, Dingkang Yang, Yuxuan Lei, Shunli Wang, Shuaibing Wang, Liuzhen Su, Kun Yang, Yuzheng Wang, Mingyang Sun, and Lihua Zhang. 2024. A Unified Self-Distillation Framework for Multimodal Sentiment Analysis with Uncertain Missing Modalities. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol. 38. 10074\u201310082. [24] Mingcheng Li, Dingkang Yang, and Lihua Zhang. 2023. Towards Robust Multimodal Sentiment Analysis under Uncertain Signal Missing. IEEE Signal Processing Letters 30 (2023), 1497\u20131501. [25] Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021). [26] Bo Liu, Li-Ming Zhan, and Xiao-Ming Wu. 2021. Contrastive Pre-training and Representation Distillation for Medical Visual Question Answering Based on Radiology Images. In MICCAI 2021. Springer International Publishing, Cham, 210\u2013220. [27] Bo Liu, Li-Ming Zhan, Xu, et al. 2021. Slake: A semantically-labeled knowledgeenhanced dataset for medical visual question answering. In 2021 ISBI. 1650\u20131654. [28] Haotian Liu, Chunyuan Li, Wu, et al. 2023. Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023). [29] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. DoRA: WeightDecomposed Low-Rank Adaptation. arXiv preprint arXiv:2402.09353 (2024). [30] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ding, et al. 2023. GPT understands, too. AI Open (2023). [31] Binh D. Nguyen, Thanh-Toan Do, Binh X Nguyen, et al. 2019. Overcoming Data Limitation in Medical Visual Question Answering. In MICCAI. Cham, 522\u2013530. [32] Alec Radford, Jong Wook Kim, Hallacy, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 8748\u20138763. [33] Sachin Mehta Sanjay Subramanian, Lucy Lu Wang et al. 2020. MedICaT: A Dataset of Medical Images, Captions, and Textual"
}