This paper describes the study of generating counterfactual pathology images for observing changes in cellular tissues associated with cancerization. Malignant lymphoma has more than 70 subtypes, and in the pathological diagnosis, a pathological image is observed to identify the subtype. Currently, there are no quantitative criteria for subtype identification, and pathologists conduct qualitative diagnoses based on their experience and intuition. Such qualitative diagnosis makes it difficult to guarantee diagnostic reproducibility, and it is desirable to develop quantitative criteria for subtype identification. Toward achieving diagnosis based on quantitative criteria, this study has tackled the generation of counterfactual pathology images using diffusion models, thereby visualizing changes in cellular tissues associated with cancerization. This paper describes the study of generating counterfactual pathology images by drawing on multiple reference papers.

This study investigates the feature representations of medical vision-language models (VLMs) to determine if medical domain-specific training truly enhances their discriminative power for downstream tasks. While numerous medical VLMs have been proposed, their feature spaces remain underexplored. We analyzed the feature distributions of representative medical and non-medical VLMs across eight diverse medical imaging datasets using dimensionality reduction and linear discriminators. Our findings reveal that medical specialization does not consistently guarantee superior image representations. Instead, non-medical VLMs employing large language models (LLMs) as text encoders demonstrate competitive or superior classification performance. Furthermore, both medical and non-medical VLMs are strongly affected by background biases, such as overlaid text strings on images, which they misinterpret as discriminative features. The results suggest that enhancing the text encoder with an LLM contributes more significantly to model performance than large-scale pre-training on noisy medical images.

Recent advances in machine learning and deep learning have accelerated research on treatment outcome prediction using medical images. In addition to prediction algorithms, the design of input factors plays a crucial role in prediction performance and model interpretability. This article reviews the input factors used for treatment outcome prediction, including patient background, clinical and laboratory information, pathological and genetic information, and medical imaging information. Particular emphasis is placed on medical imaging as a source of quantitative input factors extracted through image analysis. Image-derived features, including radiomics and deep learning-based features, are discussed from the perspective of input factor design rather than image analysis itself. Furthermore, treatment outcome prediction for extracorporeal shock wave lithotripsy (ESWL) is presented as a case study to illustrate how image-derived features and patient characteristics can be integrated into a predictive model and interpreted using feature importance analysis. Viewing medical image analysis from the perspective of input factor design provides a useful framework for improving prediction performance while enhancing the interpretability and clinical applicability of treatment outcome prediction models.

In medical image analysis, deep learning has enabled highly accurate recognition; however, several challenges remain, including the high cost of expert annotation, inter-institutional domain gaps, class imbalance, appearance variations caused by anatomical location, and ambiguous boundaries between severity classes. This article introduces our studies that leverage intrinsic data structures in medical images, such as inter-image relationships, acquisition order, anatomical location information, cluster structures, and severity ordering, as weak supervisory signals or learning constraints. Specifically, we review soft- and self-constrained clustering for group-based labeling, label-efficient representation learning using anatomical location and acquisition order, and semi-supervised domain adaptation based on cluster and ordinal structures. Through these studies, we discuss how robust computer-aided diagnosis models can be constructed from limited supervisory information by incorporating human-interpretable data structures into the learning process.