The automatic classification of pulmonary nodules is important for early cancer diagnosis. In this work, we aim to find the optimal state space model (SSM) for pulmonary nodule classification by leveraging Differentiable Architecture Search (DARTS). To achieve DARTS, we design the Differentiable State Space Supergraph (DSSS) to search for the optimal SSM. DSSS removes convolution, dimensionality expansion, and spatial attention around the SSM of Mamba while incorporating sampling-based multi-scale feature fusion. Additionally, anatomical scan provides multiple visual sequences of three-dimensional (3D) pulmonary nodules. During the supergraph search process, a knockout (KO) strategy is introduced to progressively remove low-probability candidate operations, thereby reducing the discrepancy between the continuous supergraph and the final discrete architecture while improving search efficiency. The discrete architecture obtained by DSSS is subsequently augmented with output feedback. By stacking the feedback-augmented cells, the final model, termed the Feedback-Augmented State Space Model (FASSM), is constructed. We conducted extensive experiments on the Lung Nodule Analysis 2016 (LUNA16) dataset, and the results show that the FASSM delivers outstanding performance. The architecture search was completed in less than six hours. The resulting FASSM achieved a mean classification accuracy of 93.22% across folds 5-9 with 8.59 M parameters. FASSM achieves a favorable balance between classification performance and model scale, with 8.59 M parameters and a computational cost of 20.41 GFLOPs. Related code and results have been released at: https://github.com/GuYuIMUST/DSSS-FASSM.
Electroencephalography (EEG) is a vital tool to measure and record brain activity in neuroscience and clinical applications, yet its potential is constrained by signal heterogeneity, low signal-to-noise ratios, and limited labeled datasets. In this paper, we propose FoME (Foundation Model for EEG), a novel approach using adaptive temporal-lateral attention scaling to address above-mentioned challenges. FoME is pre-trained on a diverse 1.7TB dataset of scalp and intracranial EEG recordings, comprising 745M parameters trained for 1,096k steps. Our model introduces two key innovations: a time-frequency fusion embedding technique and an adaptive temporal-lateral attention scaling (ATLAS) mechanism. These components synergistically capture complex temporal and spectral EEG dynamics, enabling FoME to adapt to varying patterns across diverse data streams and facilitate robust multi-channel modeling. Evaluations across four downstream tasks demonstrate FoME’s superior performance in classification and forecasting applications, consistently achieving state-of-the-art results. To conclude, FoME establishes a new paradigm for EEG analysis, offering a versatile foundation that advances brain-computer interfaces, clinical diagnostics, and cognitive research across neuroscience and related fields. Code will be released upon publication.
Medical lesion segmentation often depends on large, high-quality pixel-level annotations, which are costly to obtain because they require expert delineation. Recent multimodal methods use clinical text to improve visual representation learning, but they usually require paired image-report data during training and text input during inference. This dependency limits their use when paired reports are scarce or when segmentation is performed before report generation. To address this limitation, we propose TeF-SAM, a Segment Anything Model (SAM)-based framework for limited-text training and text-free medical lesion segmentation. TeF-SAM builds a Persistent Region-Informed Semantic Memory (PRISM) from the paired reports available during training. PRISM distils segmentation-relevant textual knowledge into a region-aware prototype space, allowing image features to retrieve semantic cues at inference without online text input. We further introduce Prototype-Driven Regional Lesion Detection (PRLD), which combines prototype-level semantic responses with multi-scale visual features to generate SAM-compatible prompts, including coarse masks, points, and semantic embeddings. We evaluate TeF-SAM on three lesion segmentation benchmarks spanning different anatomies and imaging modalities: QaTa-COV19, MosMedData+, and BUSI. Experimental results show that TeF-SAM achieves competitive segmentation performance using only 1%-15% of the available paired reports and requires no text input during inference. Our code will be released at: https://github.com/Jinxinya2/TeF-SAM.
Predicting the survival time of brain tumor patients remains a critical challenge in the medical prognosis. Existing methods often struggle to integrate heterogeneous data types, such as imaging and clinical records. This paper proposes a novel approach, the Three-Dimensional Feature Fusion Network (3D-FuseNet), designed to address these limitations by integrating high-dimensional imaging features with clinical variables through a late fusion strategy. 3D-FuseNet leverages artificial intelligence (AI) models to automatically extract features from 3D of four MRI sequence images and fuses them with key clinical variables to fully exploit their complementary nature. The network was evaluated using 10 widely-used optimizers and 5 machine learning (ML) methods to assess its performance comprehensively. In the best-case scenario, 3D-FuseNet achieved a mean absolute error (MAE) that is 4% lower than the best baseline model. Furthermore, it reduced the MAE by 2.7% compared to the lowest MAE obtained using individual models such as ResNet50, Unet3D encoder, and Vision Transformer. By effectively integrating multimodal data sources, 3D-FuseNet has demonstrated competitive performance in predicting survival in brain tumor patients, achieving moderate yet consistent improvements over the evaluated baseline. The proposed method outperforms existing architectures and is publicly available at: https://github.com/AIPMLab/3D-FuseNet.
Medical image segmentation often suffers from scarce training data, which limits the performance of deep learning models. Auxiliary task learning offers a promising solution by exploiting complementary supervision signals constructed solely from existing data. However, auxiliary tasks may introduce complex learning signals, and naive integration of multiple tasks often leads to limited performance gains while drastically increasing computational cost. This study presents an adaptive auxiliary task learning method to improve medical image segmentation under scarce training data. The method introduces an adaptive knowledge transfer mechanism that aggregates multi-step model updates from different tasks and model components using learnable weights, enabling more expressive knowledge transfer than prior approaches. To improve training efficiency, a task selection strategy is developed to dynamically select a small subset of beneficial tasks for each training phase. By reusing intermediate computations, it enables the evaluation of task combinations with little overhead. The proposed method is implemented using eight auxiliary tasks with diverse properties. Experiments on five public datasets demonstrate consistent improvements over existing methods, surpassing the best competing approaches by 1.47%, 0.91%, 0.46%, 0.35%, and 2.44% in Dice score. Code is available at: https://github.com/BinyanHu/AdaAuxMedSeg.
