arXiv:2607.17091v3 cs.CV · q-bio.QM Submitted 17 Jul 2026
1Department of Hematology and Stem Cell Transplantation, West German Cancer Center, NCT-West, University Hospital Essen, Essen, GER
2Institute for Artificial Intelligence in Medicine (IKIM), University Hospital Essen, Essen, GER
Diffuse gliomas carry molecular fingerprints — IDH mutation, 1p/19q codeletion, CDKN2A loss — that determine prognosis, yet routine diagnosis still hinges on morphology already encoded in the histopathology slide. This work asks whether vision can recover the genotype directly from the whole-slide image. I develop weakly-supervised, attention-based deep-learning models trained on the largest publicly available glioma WSI cohort (≈5,000 slides), pairing vision transformers with interpretable attention as explanation and cross-cohort validation as the bar for clinical credibility. The document that follows traces the path — mathematics and electronics in Hyderabad, bioinformatics on the coast of Manipal, deep learning in the Alps of Grenoble, and a PhD at Essen — from first principles to a pipeline that earns its way toward the clinic.
Keywords: computational pathology · vision transformers · attention-based multiple-instance learning · whole-slide imaging · glioma genomics · multimodal OMICS · cancer ecotypes · foundation models · interpretable AI · reproducible pipelines
A research career rarely begins where it lands. Mine started with a wide-base bachelor's in mathematics, electronics and computer science — the discipline of reasoning from first principles, and the habit of debugging circuits and code with equal patience. It moved into biology, then into machine learning, then into the specific problem of teaching a model to see what a pathologist sees and infer what a sequencer would report.
The thread across every stop has been the same: build interpretable, scalable systems that survive contact with messy real data and cross the boundary between institutions. What began as tutoring and Arduino workshops is now vision transformers on HPCs — but the goal, "to make the world a slightly better place in the slow accumulation of small things," has not moved.
Patches → features → attention → explainability. The contribution is not a single model but a reproducible one: weakly-supervised MIL that needs only slide-level labels, attention heatmaps that a pathologist can audit, and containerised pipelines (Docker / Apptainer, Snakemake) that move from one cohort — and one institution — to the next without silent failure.
Genotype-to-phenotype in adult diffuse gliomas (IDH, 1p/19q, CDKN2A). Weakly-supervised MIL, vision transformers, attention heatmaps, cross-cohort validation on HPC. Part of the CANTAR research network.
M2 thesis: Initiation of a lung adenocarcinoma cartography around TP53 activity (IAB, Dr. Cyril Boyault). IDEX & UGA Foundation scholarship. 1st place, Hackathon — Congrès National des Pharmaciens, Bordeaux.
NGS, sequence alignments, phylogenetics, functional genomics — bioinformatics proper, against the sound of the Arabian Sea.
First principles across three disciplines. Tutored matriculation students; ran a two-week Arduino / IoT workshop for ~40 participants.
Biomarker identification in esophageal squamous-cell carcinoma (ESCC). Two peer-reviewed publications.
Cloud architecture for plant breeding — GCP / AWS pipelines in Python & R.
For the poster Genetic subtype prediction in diffuse gliomas using a vision-transformer-based model (FPN 278P) — foundation models and attention-based MIL decoding molecular subtype from H&E whole-slide images. Mentors: Emre Kocakavuk, Christian Reinhardt.