Kolorektal Doku Siniflandirmasinda Kohort-Disi Model Secimi
Authors/Creators
Description
In this study, eight convolutional neural network architectures spanning 2012 to 2022 are compared on colorectal tissue classification using the official split of the PathMNIST dataset under a largely common training protocol. The test data originate from a different patient cohort than the training and validation data. Validation accuracies of all models collapse into a range of 0.36 points, whereas out-of-cohort test accuracies spread between 91.03 and 96.53 percent, and no evidence is found that the validation ranking predicts the test ranking. General accuracy and critical error rates are shown to diverge: two models indistinguishable in accuracy differ fourfold in the rate of misclassifying tumor as normal tissue. The lowest per-class performance is observed for the stroma class across all models. The findings indicate that a single aggregate metric and within-cohort validation are insufficient for model selection.
Files
58.pdf
Additional details
Additional titles
- Translated title (English)
- Out-of-Cohort Model Selection in Colorectal Tissue Classification