Dinakar Pathakota

Is it the encoder or the probe?

How much of a pathology foundation model's accuracy loss under blur, JPEG compression, and stain shift belongs to the frozen encoder, and how much to the linear probe. Across Phikon, Phikon-v2, Kaiko ViT-B/16, and DINOv2-L on NCT-CRC-HE and PatchCamelyon, a probe trained on a mix of clean and degraded embeddings recovers most of the loss — Phikon-v2 at JPEG quality 15 goes from 76.3% to 92.5% balanced accuracy — and changes which encoder looks most robust.

Paper Code

WSI inference platform

Running detection models over gigapixel slides — tumour-cell and mitotic-figure detection — covering slide ingestion, tiling, model execution, and region-level result aggregation.

Is it the encoder or the head? (CAMELYON16)

The slide-level follow-up: lymph-node metastasis detection on all 399 CAMELYON16 whole-slide images (871,867 tiles), with the same four frozen encoders and attention-based multiple-instance heads. Training the head on a mix of clean and degraded tiles lifts slide AUC under heavy JPEG compression from 63.1 to 79.7 for Phikon and from 67.9 to 81.1 for Kaiko B/16, with no cost on clean slides; it does not help Phikon-v2. The tumour signal survives in the embeddings even where the head fails, and at 95% specificity the loss falls on small metastases.

Paper Code

Vertical WSI/ML for veterinary & toxicologic pathology

A technical and market brief on preclinical, non-regulated pathology — gigapixel imaging demands comparable to the clinical setting, with far lower regulatory friction. Moving from scoping into implementation.