DIOR-ViT: Differential Ordinal Learning Vision Transformer for Cancer Classification in Pathology Images

TL;DR

DIOR-ViT addresses the ordinal nature of cancer grading by introducing differential ordinal learning into a Vision Transformer framework. It jointly optimizes categorical class prediction and pairwise differential relations between samples, using a specialized NAD loss and a weighted multi-task objective $L_{total} = L_{cat} + \lambda L_{diff}$ with $\lambda = 6.5$. Evaluated on colorectal, prostate, and gastric datasets, including a large gastric patch dataset, DIOR-ViT consistently outperforms a range of CNN and transformer baselines in accuracy, macro F1, and quadratic kappa, and ablations show the differential ordinal component and NAD loss drive the gains. The work demonstrates strong cross-domain generalization and suggests that differential ordinal learning can enhance pathology grading and potentially other ordinal-label problems in medical imaging.

Abstract

In computational pathology, cancer grading has been mainly studied as a categorical classification problem, which does not utilize the ordering nature of cancer grades such as the higher the grade is, the worse the cancer is. To incorporate the ordering relationship among cancer grades, we introduce a differential ordinal learning problem in which we define and learn the degree of difference in the categorical class labels between pairs of samples by using their differences in the feature space. To this end, we propose a transformer-based neural network that simultaneously conducts both categorical classification and differential ordinal classification for cancer grading. We also propose a tailored loss function for differential ordinal learning. Evaluating the proposed method on three different types of cancer datasets, we demonstrate that the adoption of differential ordinal learning can improve the accuracy and reliability of cancer grading, outperforming conventional cancer grading approaches. The proposed approach should be applicable to other diseases and problems as they involve ordinal relationship among class labels.

Figures (6)

• Figure 1: The overview of DIOR-ViT. DIOR-ViT consists of a feature extractor, a categorical classifier, and a differential ordinal classifier.
• Figure 2: Generation of the ground truth differential ordinal labels.
• Figure 3: Loss curves for (a) mean squared error $\mathcal{L}_{MSE}$, (b) mean absolute error $\mathcal{L}_{MAE}$, (c) ordinal cross entropy loss $\mathcal{L}_{CE_o}$, and (d) negative absolute difference log-likelihood loss $\mathcal{L}_{NAD}$. $d$ denotes the difference between the ground truth label and the predicted label in the differential ordinal classification.
• Figure 4: Comparison between Top-3 models on (a) colorectal and (b) prostate tissue datasets.
• Figure 5: Differential ordinal classification results on (a) colorectal, (b) prostate, and (c) gastric tissue datasets. The number below each image represents the output of the differential ordinal classifier.