A Controlled Decomposition of Training-Data Quantity and Class Allocation: Minority-Class Reliability in Medical Image Classification

In medical image classification, an accurate model is not the same as a trustworthy one, and in imbalanced settings the rare class is usually the one that matters. Aggregate and rank-based summaries average over a test set dominated by the majority class, so they can look acceptable while reliability on the minority class has broken. The usual study of imbalance also conflates two things: how much training data there is, and how it is allocated across classes. We separate them with a controlled decomposition on the training-data plane spanned by the minority and majority counts. Four protocols vary total quantity at balance, allocation at a fixed budget, majority exposure at a fixed minority, and minority availability at a fixed majority. Three models, from a linear baseline to a residual network, are trained from scratch under a fixed compute budget, with multiple seeds per condition as a variability estimate. Reliability is measured class-conditionally alongside the pooled summaries, and two post-hoc remedies, temperature scaling and known-prior correction, are tested against the minority gap. PatchCamelyon is the primary dataset, with ISIC 2020 as a secondary check under natural imbalance. As allocation moved from balance, pooled summaries could stay acceptable while minority recall collapsed and minority calibration deteriorated; the class split exposed the failure. Temperature scaling did not remove the class-conditional gap; known-prior correction largely closed its threshold part for the two convolutional models and over-corrected the linear one. Keywords: model calibration, class imbalance, class-conditional reliability, medical image classification, histopathology, selective prediction