Deep learning (DL) has achieved unprecedented success in precisely diagnosing dermatological lesions. However, increasing concerns over diagnostic unfairness across different demographic subgroups in DL algorithms raise issues of ethical violations and healthcare inequity. Due to limited access to the internal workings of DL algorithms, post-processing methods are widely regarded as efficient techniques for fairness enhancements in DL-based predictions. However, these methods often come at the cost of accuracy, and research exploring their application to medical images remains limited. To address these issues, we proposeFADiaFrame, an innovative post-processing framework designed to enhance both fairness and accuracy in DL-based diagnosis of dermatological lesions. Specifically, our uncertainty-aware gating calibration mechanism inFADiaFrameidentifies and calibrates untrustworthy samples, thereby enhancing diagnostic fairness, accuracy, and trustworthiness. Furthermore, we integrate this mechanism with an optimal transport method to further enhance group fairness. Extensive experiments on real-world datasets demonstrate thatFADiaFrameoutperforms existing post-processing methods in terms of both fairness and accuracy. Notably,FADiaFramepreserves diagnostic accuracy while achieving significant gains in fairness compared to pre-trained baseline models. Among all baselines, MedCLIP, withFADiaFrame, shows the most substantial improvement for the age-sensitive attribute, with accuracy increasing by 3.34% and demographic parity rising by 10.60%. Our results suggest thatFADiaFrameprovides universal applicability across diverse DL models for medical image diagnosis, ensuring both fair and accurate diagnosis across a wide range of devices and deployment contexts.

Satisfactory progress has been achieved recently in universal segmentation of CT images. Following the success of vision-language methods, there is a growing trend towards utilizing text prompts and contrastive learning to develop universal segmentation models. However, there exists a significant imbalance in information density between 3D images and text prompts. Moreover, the standard fully connected layer segmentation approach faces significant challenges with handling multiple classes and exhibits poor generalizability. To address these challenges, we propose VOxel Interacting with LAnguage method (VOILA) for universal CT image segmentation. Initially, we align voxels and language into a shared representation space and classify voxels based on cosine similarity. Subsequently, we develop the Voxel-Language Interaction framework to mitigate the impact of class imbalance caused by foreground-background discrepancies and variations in target volumes. Furthermore, a Complexity-Aware Sampling method is proposed to focus on region hard to segment, achieved by generating pseudo heatmaps from a trainable Gaussian mixture distribution. Our results indicate the proposed VOILA is capable to achieve improved performance with reduced parameters and computational cost during training. Furthermore, it demonstrates significant generalizability across diverse datasets without additional fine-tuning.

Accurate prediction of pathologic complete response to neoadjuvant chemotherapy non-invasively before treatment via dynamic contrast-enhanced magnetic resonance imaging is vital for developing a personalized therapy strategy. However, the application of deep learning in this domain is characterized by its black-box nature, largely relying on post-hoc analysis to interpret final decision-making. This reliance results in a lack of self-interpretability in the operational mechanisms of feature extraction, feature fusion, and decision-making. Moreover, these models have demonstrated unsatisfactory prediction performance due to insufficient feature modeling. To address these issues, we propose a self-interpretable deep learning network that can provide the intrinsic interpretability of feature extraction, multi-scale feature fusion, and final prediction. First, the interpretable perception module is designed to extract features both effectively and interpretably. Furthermore, the interpretable adaptive multi-scale feature fusion module is proposed to fuse multi-scale features. Finally, an end-to-end self-interpretable deep learning network is presented to predict pathologic complete response with self-interpretability. Validated on a multi-center pre-treatment dynamic contrast-enhanced magnetic resonance imaging dataset, our self-interpretable deep learning network outperforms state-of-the-art methods in both prediction performance and self-interpretability, improving the area under the receiver operating characteristic curve by at least 4.81% while providing both qualitative and quantitative self-interpretability. Our study demonstrates that our proposed self-interpretable deep learning network can extract key information from pre-treatment breast dynamic contrast-enhanced magnetic resonance imaging while enhancing both the prediction performance and the transparency of the model, thereby improving its trustworthiness in clinical settings.