Abstract:Endorectal ultrasound (ERUS) is an important imaging modality that provides high reliability for diagnosing the depth and boundary of invasion in colorectal cancer. However, the lack of a large-scale ERUS dataset with high-quality annotations hinders the development of automatic ultrasound diagnostics. In this paper, we collected and annotated the first benchmark dataset that covers diverse ERUS scenarios, i.e. colorectal cancer segmentation, detection, and infiltration depth staging. Our ERUS-10K dataset comprises 77 videos and 10,000 high-resolution annotated frames. Based on this dataset, we further introduce a benchmark model for colorectal cancer segmentation, named the Adaptive Sparse-context TRansformer (ASTR). ASTR is designed based on three considerations: scanning mode discrepancy, temporal information, and low computational complexity. For generalizing to different scanning modes, the adaptive scanning-mode augmentation is proposed to convert between raw sector images and linear scan ones. For mining temporal information, the sparse-context transformer is incorporated to integrate inter-frame local and global features. For reducing computational complexity, the sparse-context block is introduced to extract contextual features from auxiliary frames. Finally, on the benchmark dataset, the proposed ASTR model achieves a 77.6% Dice score in rectal cancer segmentation, largely outperforming previous state-of-the-art methods.
Abstract:Aside from offering state-of-the-art performance in medical image generation, denoising diffusion probabilistic models (DPM) can also serve as a representation learner to capture semantic information and potentially be used as an image representation for downstream tasks, e.g., segmentation. However, these latent semantic representations rely heavily on labor-intensive pixel-level annotations as supervision, limiting the usability of DPM in medical image segmentation. To address this limitation, we propose an enhanced diffusion segmentation model, called TextDiff, that improves semantic representation through inexpensive medical text annotations, thereby explicitly establishing semantic representation and language correspondence for diffusion models. Concretely, TextDiff extracts intermediate activations of the Markov step of the reverse diffusion process in a pretrained diffusion model on large-scale natural images and learns additional expert knowledge by combining them with complementary and readily available diagnostic text information. TextDiff freezes the dual-branch multi-modal structure and mines the latent alignment of semantic features in diffusion models with diagnostic descriptions by only training the cross-attention mechanism and pixel classifier, making it possible to enhance semantic representation with inexpensive text. Extensive experiments on public QaTa-COVID19 and MoNuSeg datasets show that our TextDiff is significantly superior to the state-of-the-art multi-modal segmentation methods with only a few training samples.
Abstract:Black-box tuning has attracted recent attention due to that the structure or inner parameters of advanced proprietary models are not accessible. Proxy-tuning provides a test-time output adjustment for tuning black-box language models. It applies the difference of the output logits before and after tuning a smaller white-box "proxy" model to improve the black-box model. However, this technique serves only as a decoding-time algorithm, leading to an inconsistency between training and testing which potentially limits overall performance. To address this problem, we introduce Consistent Proxy Tuning (CPT), a simple yet effective black-box tuning method. Different from Proxy-tuning, CPT additionally exploits the frozen large black-box model and another frozen small white-box model, ensuring consistency between training-stage optimization objective and test-time proxies. This consistency benefits Proxy-tuning and enhances model performance. Note that our method focuses solely on logit-level computation, which makes it model-agnostic and applicable to any task involving logit classification. Extensive experimental results demonstrate the superiority of our CPT in both black-box tuning of Large Language Models (LLMs) and Vision-Language Models (VLMs) across various datasets. The code is available at https://github.com/chunmeifeng/CPT.
Abstract:Albeit progress has been made in Composed Image Retrieval (CIR), we empirically find that a certain percentage of failure retrieval results are not consistent with their relative captions. To address this issue, this work provides a Visual Question Answering (VQA) perspective to boost the performance of CIR. The resulting VQA4CIR is a post-processing approach and can be directly plugged into existing CIR methods. Given the top-C retrieved images by a CIR method, VQA4CIR aims to decrease the adverse effect of the failure retrieval results being inconsistent with the relative caption. To find the retrieved images inconsistent with the relative caption, we resort to the "QA generation to VQA" self-verification pipeline. For QA generation, we suggest fine-tuning LLM (e.g., LLaMA) to generate several pairs of questions and answers from each relative caption. We then fine-tune LVLM (e.g., LLaVA) to obtain the VQA model. By feeding the retrieved image and question to the VQA model, one can find the images inconsistent with relative caption when the answer by VQA is inconsistent with the answer in the QA pair. Consequently, the CIR performance can be boosted by modifying the ranks of inconsistently retrieved images. Experimental results show that our proposed method outperforms state-of-the-art CIR methods on the CIRR and Fashion-IQ datasets.
Abstract:3D Visual Grounding (3DVG) aims at localizing 3D object based on textual descriptions. Conventional supervised methods for 3DVG often necessitate extensive annotations and a predefined vocabulary, which can be restrictive. To address this issue, we propose a novel visual programming approach for zero-shot open-vocabulary 3DVG, leveraging the capabilities of large language models (LLMs). Our approach begins with a unique dialog-based method, engaging with LLMs to establish a foundational understanding of zero-shot 3DVG. Building on this, we design a visual program that consists of three types of modules, i.e., view-independent, view-dependent, and functional modules. These modules, specifically tailored for 3D scenarios, work collaboratively to perform complex reasoning and inference. Furthermore, we develop an innovative language-object correlation module to extend the scope of existing 3D object detectors into open-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot approach can outperform some supervised baselines, marking a significant stride towards effective 3DVG.
Abstract:Composed image retrieval (CIR) is the task of retrieving specific images by using a query that involves both a reference image and a relative caption. Most existing CIR models adopt the late-fusion strategy to combine visual and language features. Besides, several approaches have also been suggested to generate a pseudo-word token from the reference image, which is further integrated into the relative caption for CIR. However, these pseudo-word-based prompting methods have limitations when target image encompasses complex changes on reference image, e.g., object removal and attribute modification. In this work, we demonstrate that learning an appropriate sentence-level prompt for the relative caption (SPRC) is sufficient for achieving effective composed image retrieval. Instead of relying on pseudo-word-based prompts, we propose to leverage pretrained V-L models, e.g., BLIP-2, to generate sentence-level prompts. By concatenating the learned sentence-level prompt with the relative caption, one can readily use existing text-based image retrieval models to enhance CIR performance. Furthermore, we introduce both image-text contrastive loss and text prompt alignment loss to enforce the learning of suitable sentence-level prompts. Experiments show that our proposed method performs favorably against the state-of-the-art CIR methods on the Fashion-IQ and CIRR datasets. The source code and pretrained model are publicly available at https://github.com/chunmeifeng/SPRC
Abstract:Universal Domain Adaptation (UniDA) aims to distinguish common and private classes between the source and target domains where domain shift exists. Recently, due to more stringent data restrictions, researchers have introduced Source-Free UniDA (SF-UniDA) in more realistic scenarios. SF-UniDA methods eliminate the need for direct access to source samples when performing adaptation to the target domain. However, existing SF-UniDA methods still require an extensive quantity of labeled source samples to train a source model, resulting in significant labeling costs. To tackle this issue, we present a novel Classifier-Oriented Calibration (COCA) method. This method, which leverages textual prototypes, is formulated for the source model based on few-shot learning. Specifically, we propose studying few-shot learning, usually explored for closed-set scenarios, to identify common and domain-private classes despite a significant domain shift between source and target domains. Essentially, we present a novel paradigm based on the vision-language model to learn SF-UniDA and hugely reduce the labeling costs on the source domain. Experimental results demonstrate that our approach outperforms state-of-the-art UniDA and SF-UniDA models.
Abstract:While multi-modal learning has been widely used for MRI reconstruction, it relies on paired multi-modal data which is difficult to acquire in real clinical scenarios. Especially in the federated setting, the common situation is that several medical institutions only have single-modal data, termed the modality missing issue. Therefore, it is infeasible to deploy a standard federated learning framework in such conditions. In this paper, we propose a novel communication-efficient federated learning framework, namely Fed-PMG, to address the missing modality challenge in federated multi-modal MRI reconstruction. Specifically, we utilize a pseudo modality generation mechanism to recover the missing modality for each single-modal client by sharing the distribution information of the amplitude spectrum in frequency space. However, the step of sharing the original amplitude spectrum leads to heavy communication costs. To reduce the communication cost, we introduce a clustering scheme to project the set of amplitude spectrum into finite cluster centroids, and share them among the clients. With such an elaborate design, our approach can effectively complete the missing modality within an acceptable communication cost. Extensive experiments demonstrate that our proposed method can attain similar performance with the ideal scenario, i.e., all clients have the full set of modalities. The source code will be released.
Abstract:Federated Learning (FL) enables multiple clients to collaboratively learn in a distributed way, allowing for privacy protection. However, the real-world non-IID data will lead to client drift which degrades the performance of FL. Interestingly, we find that the difference in logits between the local and global models increases as the model is continuously updated, thus seriously deteriorating FL performance. This is mainly due to catastrophic forgetting caused by data heterogeneity between clients. To alleviate this problem, we propose a new algorithm, named FedCSD, a Class prototype Similarity Distillation in a federated framework to align the local and global models. FedCSD does not simply transfer global knowledge to local clients, as an undertrained global model cannot provide reliable knowledge, i.e., class similarity information, and its wrong soft labels will mislead the optimization of local models. Concretely, FedCSD introduces a class prototype similarity distillation to align the local logits with the refined global logits that are weighted by the similarity between local logits and the global prototype. To enhance the quality of global logits, FedCSD adopts an adaptive mask to filter out the terrible soft labels of the global models, thereby preventing them to mislead local optimization. Extensive experiments demonstrate the superiority of our method over the state-of-the-art federated learning approaches in various heterogeneous settings. The source code will be released.
Abstract:Benefiting from prompt tuning, recent years have witnessed the promising performance of pre-trained vision-language models, e.g., CLIP, on versatile downstream tasks. In this paper, we focus on a particular setting of learning adaptive prompts on the fly for each test sample from an unseen new domain, which is known as test-time prompt tuning (TPT). Existing TPT methods typically rely on data augmentation and confidence selection. However, conventional data augmentation techniques, e.g., random resized crops, suffers from the lack of data diversity, while entropy-based confidence selection alone is not sufficient to guarantee prediction fidelity. To address these issues, we propose a novel TPT method, named DiffTPT, which leverages pre-trained diffusion models to generate diverse and informative new data. Specifically, we incorporate augmented data by both conventional method and pre-trained stable diffusion to exploit their respective merits, improving the models ability to adapt to unknown new test data. Moreover, to ensure the prediction fidelity of generated data, we introduce a cosine similarity-based filtration technique to select the generated data with higher similarity to the single test sample. Our experiments on test datasets with distribution shifts and unseen categories demonstrate that DiffTPT improves the zero-shot accuracy by an average of 5.13\% compared to the state-of-the-art TPT method. Our code and models will be publicly released.