Limited computational resources make full fine-tuning increasingly difficult as vision models grow, while digital pathology also requires reliable adaptation to new tissue domains. This paper presents a parameter-efficient colorectal histopathology classifier based on a pretrained Data-efficient Image Transformer (DeiT-Tiny) and low-rank adaptation (LoRA). The model was evaluated on PathMNIST, a nine-class collection of hematoxylin-and-eosin colorectal tissue patches with a predefined 7,180-image test set originating from a different clinical center. LoRA modules were inserted into the query and value projections of each self-attention block, while the pretrained backbone was frozen and the classification head remained trainable. Only 149,193 of 5,675,346 parameters (2.63%) were optimized. After ten epochs, best validation accuracy reached 96.61%. On the test set, the model obtained 93.30% accuracy, 0.9335 weighted precision, 0.9320 weighted F1-score, 0.9917 mean specificity, and 0.9951 macro one-vs-rest ROC AUC. Error analysis showed strong recognition of adipose, background, lymphocytes, mucus, and tumor epithelium, whereas debris and cancer-associated stroma remained more difficult. The results establish a reproducible parameter-efficient baseline for the third objective of a broader data-scarce pathology framework. However, the present experiment measures parameter efficiency rather than few-shot label efficiency; controlled label-fraction, full-fine-tuning, calibration, and out-of-distribution studies are required before clinical interpretation.
 

digital pathology, histopathology, low-rank adaptation, parameter-efficient fine-tuning, Vision Transformer, DeiT, PathMNIST.