Towards a Comprehensive, Efficient and Promptable Anatomic Structure Segmentation Model using 3D Whole-body CT Scans

Segment anything model (SAM) demonstrates strong generalization ability on natural image segmentation. However, its direct adaption in medical image segmentation tasks shows significant performance drops with inferior accuracy and unstable results. It may also requires an excessive number of prompt points to obtain a reasonable accuracy. For segmenting 3D radiological CT or MRI scans, a 2D SAM model has to separately handle hundreds of 2D slices. Although quite a few studies explore adapting SAM into medical image volumes, the efficiency of 2D adaption methods is unsatisfactory and 3D adaptation methods only capable of segmenting specific organs/tumors. In this work, we propose a comprehensive and scalable 3D SAM model for whole-body CT segmentation, named CT-SAM3D. Instead of adapting SAM, we propose a 3D promptable segmentation model using a (nearly) fully labeled CT dataset. To train CT-SAM3D effectively, ensuring the model's accurate responses to higher-dimensional spatial prompts is crucial, and 3D patch-wise training is required due to GPU memory constraints. For this purpose, we propose two key technical developments: 1) a progressively and spatially aligned prompt encoding method to effectively encode click prompts in local 3D space; and 2) a cross-patch prompt learning scheme to capture more 3D spatial context, which is beneficial for reducing the editing workloads when interactively prompting on large organs. CT-SAM3D is trained and validated using a curated dataset of 1204 CT scans containing 107 whole-body anatomies, reporting significantly better quantitative performance against all previous SAM-derived models by a large margin with much fewer click prompts. Our model can handle segmenting unseen organ as well. Code, data, and our 3D interactive segmentation tool with quasi-real-time responses will be made publicly available.

Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration

Establishing dense anatomical correspondence across distinct imaging modalities is a foundational yet challenging procedure for numerous medical image analysis studies and image-guided radiotherapy. Existing multi-modality image registration algorithms rely on statistical-based similarity measures or local structural image representations. However, the former is sensitive to locally varying noise, while the latter is not discriminative enough to cope with complex anatomical structures in multimodal scans, causing ambiguity in determining the anatomical correspondence across scans with different modalities. In this paper, we propose a modality-agnostic structural representation learning method, which leverages Deep Neighbourhood Self-similarity (DNS) and anatomy-aware contrastive learning to learn discriminative and contrast-invariance deep structural image representations (DSIR) without the need for anatomical delineations or pre-aligned training images. We evaluate our method on multiphase CT, abdomen MR-CT, and brain MR T1w-T2w registration. Comprehensive results demonstrate that our method is superior to the conventional local structural representation and statistical-based similarity measures in terms of discriminability and accuracy.

Model Tailor: Mitigating Catastrophic Forgetting in Multi-modal Large Language Models

Catastrophic forgetting emerges as a critical challenge when fine-tuning multi-modal large language models (MLLMs), where improving performance on unseen tasks often leads to a significant performance drop on the original tasks. This paper presents a comprehensive analysis of catastrophic forgetting in MLLMs and introduces a post-training adjustment method called Model Tailor. Our method primarily preserves the pre-trained parameters while replacing a small number ($\leq$ 10\%) of fine-tuned parameters, maintaining $\sim$ 99\% effectiveness on original tasks versus pre-training, and achieving $\sim$ 97\% on new tasks compared to standard fine-tuning. Specifically, we derive a sparse mask to identify the "model patch", based on a fusion strategy that integrates salience and sensitivity analysis. Subsequently, a compensation mechanism is introduced to "decorate the patch", enhancing the model's performance on both target and original tasks. Additionally, our method is adaptable to multi-task scenarios. Through extensive experiments on InstructBLIP and LLaVA-1.5 in both image captioning and visual question answering tasks, our approach demonstrates significant task adaptability while preserving inherent pre-trained capabilities.

VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding

Large-scale pre-trained models have achieved remarkable success in various computer vision tasks. A standard approach to leverage these models is to fine-tune all model parameters for downstream tasks, which poses challenges in terms of computational and storage costs. Recently, inspired by Natural Language Processing (NLP), parameter-efficient transfer learning has been successfully applied to vision tasks. However, most existing techniques primarily focus on single-task adaptation, and despite limited research on multi-task adaptation, these methods often exhibit suboptimal training and inference efficiency. In this paper, we first propose an once-for-all Vision Multi-Task Adapter (VMT-Adapter), which strikes approximately O(1) training and inference efficiency w.r.t task number. Concretely, VMT-Adapter shares the knowledge from multiple tasks to enhance cross-task interaction while preserves task-specific knowledge via independent knowledge extraction modules. Notably, since task-specific modules require few parameters, VMT-Adapter can handle an arbitrary number of tasks with a negligible increase of trainable parameters. We also propose VMT-Adapter-Lite, which further reduces the trainable parameters by learning shared parameters between down- and up-projections. Extensive experiments on four dense scene understanding tasks demonstrate the superiority of VMT-Adapter(-Lite), achieving a 3.96%(1.34%) relative improvement compared to single-task full fine-tuning, while utilizing merely ~1% (0.36%) trainable parameters of the pre-trained model.

MmAP : Multi-modal Alignment Prompt for Cross-domain Multi-task Learning

Multi-Task Learning (MTL) is designed to train multiple correlated tasks simultaneously, thereby enhancing the performance of individual tasks. Typically, a multi-task network structure consists of a shared backbone and task-specific decoders. However, the complexity of the decoders increases with the number of tasks. To tackle this challenge, we integrate the decoder-free vision-language model CLIP, which exhibits robust zero-shot generalization capability. Recently, parameter-efficient transfer learning methods have been extensively explored with CLIP for adapting to downstream tasks, where prompt tuning showcases strong potential. Nevertheless, these methods solely fine-tune a single modality (text or visual), disrupting the modality structure of CLIP. In this paper, we first propose Multi-modal Alignment Prompt (MmAP) for CLIP, which aligns text and visual modalities during fine-tuning process. Building upon MmAP, we develop an innovative multi-task prompt learning framework. On the one hand, to maximize the complementarity of tasks with high similarity, we utilize a gradient-driven task grouping method that partitions tasks into several disjoint groups and assign a group-shared MmAP to each group. On the other hand, to preserve the unique characteristics of each task, we assign an task-specific MmAP to each task. Comprehensive experiments on two large multi-task learning datasets demonstrate that our method achieves significant performance improvements compared to full fine-tuning while only utilizing approximately 0.09% of trainable parameters.

SAMv2: A Unified Framework for Learning Appearance, Semantic and Cross-Modality Anatomical Embeddings

Identifying anatomical structures (e.g., lesions or landmarks) in medical images plays a fundamental role in medical image analysis. As an exemplar-based landmark detection method, Self-supervised Anatomical eMbedding (SAM) learns a discriminative embedding for each voxel in the image and has shown promising results on various tasks. However, SAM still faces challenges in: (1) differentiating voxels with similar appearance but different semantic meanings (\textit{e.g.}, two adjacent structures without clear borders); (2) matching voxels with similar semantics but markedly different appearance (e.g., the same vessel before and after contrast injection); and (3) cross-modality matching (e.g., CT-MRI registration). To overcome these challenges, we propose SAMv2, which is a unified framework designed to learn appearance, semantic, and cross-modality anatomical embeddings. Specifically, SAMv2 incorporates three key innovations: (1) semantic embedding learning with prototypical contrastive loss; (2) a fixed-point-based matching strategy; and (3) an iterative approach for cross-modality embedding learning. We thoroughly evaluated SAMv2 across three tasks, including one-shot landmark detection, lesion tracking on longitudinal CT scans, and CT-MRI affine/rigid registration with varying field of view. Our results suggest that SAMv2 outperforms SAM and other state-of-the-art methods, offering a robust and versatile approach for landmark based medical image analysis tasks. Code and trained models are available at: https://github.com/alibaba-damo-academy/self-supervised-anatomical-embedding-v2

SAME++: A Self-supervised Anatomical eMbeddings Enhanced medical image registration framework using stable sampling and regularized transformation

Image registration is a fundamental medical image analysis task. Ideally, registration should focus on aligning semantically corresponding voxels, i.e., the same anatomical locations. However, existing methods often optimize similarity measures computed directly on intensities or on hand-crafted features, which lack anatomical semantic information. These similarity measures may lead to sub-optimal solutions where large deformations, complex anatomical differences, or cross-modality imagery exist. In this work, we introduce a fast and accurate method for unsupervised 3D medical image registration building on top of a Self-supervised Anatomical eMbedding (SAM) algorithm, which is capable of computing dense anatomical correspondences between two images at the voxel level. We name our approach SAM-Enhanced registration (SAME++), which decomposes image registration into four steps: affine transformation, coarse deformation, deep non-parametric transformation, and instance optimization. Using SAM embeddings, we enhance these steps by finding more coherent correspondence and providing features with better semantic guidance. We extensively evaluated SAME++ using more than 50 labeled organs on three challenging inter-subject registration tasks of different body parts. As a complete registration framework, SAME++ markedly outperforms leading methods by $4.2\%$ - $8.2\%$ in terms of Dice score while being orders of magnitude faster than numerical optimization-based methods. Code is available at \url{https://github.com/alibaba-damo-academy/same}.

Combining Past, Present and Future: A Self-Supervised Approach for Class Incremental Learning

Class Incremental Learning (CIL) aims to handle the scenario where data of novel classes occur continuously and sequentially. The model should recognize the sequential novel classes while alleviating the catastrophic forgetting. In the self-supervised manner, it becomes more challenging to avoid the conflict between the feature embedding spaces of novel classes and old ones without any class labels. To address the problem, we propose a self-supervised CIL framework CPPF, meaning Combining Past, Present and Future. In detail, CPPF consists of a prototype clustering module (PC), an embedding space reserving module (ESR) and a multi-teacher distillation module (MTD). 1) The PC and the ESR modules reserve embedding space for subsequent phases at the prototype level and the feature level respectively to prepare for knowledge learned in the future. 2) The MTD module maintains the representations of the current phase without the interference of past knowledge. One of the teacher networks retains the representations of the past phases, and the other teacher network distills relation information of the current phase to the student network. Extensive experiments on CIFAR100 and ImageNet100 datasets demonstrate that our proposed method boosts the performance of self-supervised class incremental learning. We will release code in the near future.

HODN: Disentangling Human-Object Feature for HOI Detection

The task of Human-Object Interaction (HOI) detection is to detect humans and their interactions with surrounding objects, where transformer-based methods show dominant advances currently. However, these methods ignore the relationship among humans, objects, and interactions: 1) human features are more contributive than object ones to interaction prediction; 2) interactive information disturbs the detection of objects but helps human detection. In this paper, we propose a Human and Object Disentangling Network (HODN) to model the HOI relationships explicitly, where humans and objects are first detected by two disentangling decoders independently and then processed by an interaction decoder. Considering that human features are more contributive to interaction, we propose a Human-Guide Linking method to make sure the interaction decoder focuses on the human-centric regions with human features as the positional embeddings. To handle the opposite influences of interactions on humans and objects, we propose a Stop-Gradient Mechanism to stop interaction gradients from optimizing the object detection but to allow them to optimize the human detection. Our proposed method achieves competitive performance on both the V-COCO and the HICO-Det datasets. It can be combined with existing methods easily for state-of-the-art results.

Seeing in Flowing: Adapting CLIP for Action Recognition with Motion Prompts Learning

The Contrastive Language-Image Pre-training (CLIP) has recently shown remarkable generalization on "zero-shot" training and has applied to many downstream tasks. We explore the adaptation of CLIP to achieve a more efficient and generalized action recognition method. We propose that the key lies in explicitly modeling the motion cues flowing in video frames. To that end, we design a two-stream motion modeling block to capture motion and spatial information at the same time. And then, the obtained motion cues are utilized to drive a dynamic prompts learner to generate motion-aware prompts, which contain much semantic information concerning human actions. In addition, we propose a multimodal communication block to achieve a collaborative learning and further improve the performance. We conduct extensive experiments on HMDB-51, UCF-101, and Kinetics-400 datasets. Our method outperforms most existing state-of-the-art methods by a significant margin on "few-shot" and "zero-shot" training. We also achieve competitive performance on "closed-set" training with extremely few trainable parameters and additional computational costs.