Deep learning offers notable promise for computational pathology, but its performance is constrained by the need for extensively annotated datasets, which are costly and laborious to produce. Self-Supervised Learning (SSL) provides an effective paradigm for learning discriminative representations from unannotated pathological images. However, existing SSL methods often overlook domain-specific characteristics of pathological images and suffer from the adverse effects of low-quality negative samples, leading to sub-optimal feature representations for downstream tasks. To overcome these limitations, we propose a novel Domain-Specific Self-supervised Contrastive Learning (DSSCL) framework, which incorporates two novel components: (1) a Stain-Separation Based Data Augmentation (SSDA) module that enhances stain-aware representation learning by fusing stain-separated components with original hematoxylin and eosin images, and (2) a Contrast-Aware Pair Refinement (CAPR) module that improves feature discriminability by filtering potential positives and mining hard negatives, thereby mitigating the influence of low-quality negatives. Extensive experiments demonstrate that DSSCL achieves comparable accuracy in classification tasks using only 0.1% labeled data compared to a network fine-tuned from ImageNet with 10% labeled data, while also delivering competitive performance in detection and segmentation tasks, underscoring its effectiveness in learning transferable and robust feature representations across diverse downstream tasks. The code is available at https://github.com/junjianli106/DSSCL.

Accurate prediction of pathological complete response (pCR) is useful for clinical precision treatment of lung cancer. However, most existing pCR prediction methods are based on either convolutional neural networks (CNNs) or Transformers, which cannot effectively capture 3D global information or fuse features. Therefore, this study proposes a novel multi-scale CNN-Mamba network (MSCM-Net) to achieve accurate pCR prediction on CT scans of lung cancer patients. In each stage of the hybrid encoder, after channel splitting for parameter reduction, we design the CNN branch and Mamba branch to extract local and global features. Specifically, in the Mamba branch, we propose a novel intra-slice and inter-slice scanning mechanism to implement 8-way 3D scanning, thereby effectively capturing 3D global information. Furthermore, to better fuse CNN and Mamba features, we design a novel multi-scale aware feature fusion module with channel-level and multi-scale spatial level fusion. The proposed method is evaluated on a private dataset including 108 lung cancer patients who underwent neoadjuvant chemoimmunotherapy. Experimental results demonstrate that MSCM-Net achieves the best accuracy of 83.33% and an area under curve of 84.08%. Furthermore, results on the esophageal cancer and stroke prognosis prediction tasks validate the generalizability of MSCM-Net.

Accurately predicting the survival of patients with esophageal cancer after esophagectomy is crucial for clinical precision treatment. However, the existing methods of predicting Overall Survival time (OStime) mostly build supervised learning with the uncensored data, ignoring the potential information hidden in the censored data. To utilize the information hidden in the clinically abundant censored data, we propose a Semi-Supervised Learning with Adaptive pseudo-label Selection and Correction (SSLASC) to predict the OStime of esophageal cancer using both uncensored and censored data. Specifically, we first transform the OStime regression problem to a classification task followed by Softmax Expected Value Refinement (SEVR) and train a Transformer network using the uncensored data, which is then used to predict the OStime for the censored data. Secondly, we design an adaptive pseudo-label selection strategy to dynamically select more classes and more balanced samples from the predicted censored data by allocating adaptive thresholds for different classes of samples when performing pseudo-label selection. Finally, a distribution correction and a meta label correction modules are proposed to make the selected pseudo-labels closer to the real overall OStime. We test SSLASC on an internal dataset and two external datasets with sample sizes of 327, 104, and 16, respectively. The experimental results demonstrate that SSLASC achieves Mean Absolute Error (MAE) of 12.23, 12.64, and 12.47 months on the three test datasets. Compared to the optimal State-Of-The-Art (SOTA) method, SSLASC improves performance by 1.09, 1.07, and 1.09 months, respectively. In addition, SSLASC also achieves the best performance in dichotomized survival analysis.

A novel hybrid model combining a convolutional neural network (CNN) and a low-complexity Transformer network is introduced for predicting lung cancer response to neoadjuvant chemoimmunotherapy using computed tomography scans. This approach is crucial as it assists clinicians in identifying patients likely to benefit from treatment and in assessing their prognosis. The model employs channel splitting to minimize parameter count. It then leverages both CNN for local feature extraction and a streamlined Transformer for global feature comprehension. To enhance efficiency, a novel self-attention mechanism is implemented, focusing on feature aggregation and element-wise multiplication. To address the different semantic meanings of features, an attention-based module is designed to seamlessly integrate features from both networks, employing a process of coarse fusion, attention computation, and fine fusion. When evaluated with data from 232 lung cancer patients who have undergone neoadjuvant chemoimmunotherapy, the model demonstrates exceptional performance, achieving a Dice score of 47.04% and a 95.00% Hausdorff distance of 25.12 mm, outperforming existing methods. Additionally, it has only 2.91×10⁶ parameters and 52.95×10⁹ floating point operations. Moreover, the model’s predictive accuracy in tumor diameter estimation is beneficial for treatment planning. Its robustness is further validated through its application in stroke lesion prediction, indicating its broad applicability.

Predicting mortality risk in the Intensive Care Unit (ICU) using Electronic Medical Records (EMR) is crucial for identifying patients in need of immediate attention. However, the incompleteness and the variability of EMR features for each patient make mortality prediction challenging. This study proposes a multimodal representation learning framework based on a novel personalized graph-based fusion approach to address these challenges. The proposed approach involves constructing patient-specific modality aggregation graphs to provide information about the features associated with each patient from incomplete multimodal data, enabling the effective and explainable fusion of the incomplete features. Modality-specific encoders are employed to encode each modality feature separately. To tackle the variability and incompleteness of input features among patients, a novel personalized graph-based fusion method is proposed to fuse patient-specific multimodal feature representations based on the constructed modality aggregation graphs. Furthermore, a MultiModal Gated Contrastive Representation Learning (MMGCRL) method is proposed to facilitate capturing adequate complementary information from multimodal representations and improve model performance. We evaluate the proposed framework using the large-scale ICU dataset, MIMIC-III. Experimental results demonstrate its effectiveness in mortality prediction, outperforming several state-of-the-art methods.

Glioma survival risk prediction is of great significance for the individualized treatment and assessment programs. Currently, most deep learning based survival prediction paradigms rely on invasive and expensive histopathology and genomics methods. However, magnetic resonance imaging (MRI) has emerged as a promising non-invasive alternative with significant prognostic potential. To leverage the benefits of MRI, we propose a segmentation-guided fully automated multimodal MRI-based survival network (SGS-Net), which can simultaneously perform glioma segmentation and survival risk prediction. Specifically, the task interrelation is addressed using a hybrid convolutional neural network-Transformer (CNN-Transformer) encoder to represent the shared high-level semantic features by co-training a decoder for glioma segmentation and a Cox model for survival prediction. Then, to ensure the effective representation of the high-level features, glioma segmentation as an auxiliary task is utilized to guide survival prediction by jointly optimizing the segmentation loss and the Cox partial log-likelihood loss. Furthermore, a pair-wise ranking loss is designed to allow the network to learn the survival difference between patients. To balance the multi-task losses, an uncertain weight manner is adopted to adaptively adjust the weights for preventing task bias. Finally, the proposed SGS-Net is assessed using a publicly available multi-institutional dataset. Experimental and visual results show that SGS-Net achieves promising segmentation performance and obtains a C-index of 81.07% for survival risk prediction, which outperforms several existing state-of-the-art methods and even histopathology-based methods. In addition, Kaplan-Meier survival analysis confirms that the prognosis risk generated by SGS-Net is consistent with the prior prognosis based on the grading or genotyping paradigms.