Master's Thesis: Fine-Tuned Foundation Models for Internal CLL Cluster Segmentation

My master's thesis builds and evaluates a pipeline for automatically segmenting the internal cell clusters inside CLL (chronic lymphocytic leukemia) spheroids from brightfield microscopy time-lapse. These clusters are what actually carry the drug-response signal: when a drug works, they disperse and the spheroid shrinks. They're also low-contrast and nearly invisible to off-the-shelf models, and only 91 of roughly 97,000 frames in the archive have expert annotations.

The main result runs against what you'd expect. I fine-tuned MedSAM2, a video-capable foundation segmentation model, and compared it to a plain supervised U-Net. With oracle prompts (points and boxes taken directly from the ground-truth mask), MedSAM2 wins, 0.72 vs. 0.67 Pixel IoU. But oracle prompts don't exist at deployment. On an unlabeled archive, the prompt has to be generated automatically from the image itself. Once I switched to that realistic setting, the ranking flipped completely: the plain U-Net beat fine-tuned MedSAM2 on every single experiment, 0.62 vs. 0.56 Pixel IoU. The foundation model's advantage only held under conditions that don't exist in production.

I also tested whether SAM2's video memory helps track clusters over time. It improved frame-to-frame consistency but made final-frame accuracy worse, since its memory is built for continuous video, not sparse, multi-hour microscopy captures.

For the biological payoff, I ran the deployed U-Net across a 23-experiment cohort and showed that image-derived cluster count tracks real cell viability, controlling for batch differences between experiments with a mixed-effects model. It generalizes to experiments the model never trained on.

The most useful finding is probably the negative one. I tried five different ways to work around the annotation bottleneck (pseudo-labeling, self-supervised pretraining, expert-in-the-loop relabeling, and more), and almost none of them moved the accuracy needle. That points to the real constraint being the amount and quality of annotated data, not model architecture. It's a less flattering conclusion than "the foundation model wins," but a more honest and useful one for anyone actually deploying this.