Research
Current research, plus two DOI-archived papers
The featured project is active right now — methodology and scope are locked in, results are not, and that's stated plainly below.
Semi-Supervised Melanoma Nuclei Segmentation for Computer-Aided Diagnosis
Melanoma is diagnosed by pathologists manually examining histopathology slides — can a model learn to segment the relevant cell nuclei from only a handful of labeled examples?
Abstract
Expert-labeled medical images are expensive and scarce in practice, which makes fully supervised nuclei segmentation for melanoma histopathology hard to scale. This project tests whether semi-supervised learning can close that gap. Working with the PUMA histopathology dataset (205 annotated images), a U-Net segmentation baseline reached 0.91 Dice under full supervision — within 1.2 points of the 0.916 Dice reported by the benchmark study this work builds on (Akbarpour et al., J. Imaging 11(8):274, 2025). Four semi-supervised techniques — Mean Teacher consistency regularization, a semi-supervised GAN, autoencoder pretraining, and confidence-based pseudo-labeling — were then rigorously compared against a matched supervised baseline under simulated label scarcity (15 and 40 labeled images). The manuscript is currently being written.
Contributions
- A U-Net supervised baseline on the PUMA dataset reaching 0.91 Dice, closely matching the benchmark study it's compared against
- A controlled comparison of four distinct semi-supervised techniques — Mean Teacher, semi-supervised GAN, autoencoder pretraining, and pseudo-labeling — against the same supervised baseline under simulated label scarcity
- A demonstration that GAN-based semi-supervision reaches near-parity with full supervision using only 15 labeled images
- A follow-up domain-diversity experiment using real out-of-domain histopathology data sourced from the public TCGA-SKCM archive via the NIH GDC API
Methodology
U-Net segmentation baseline trained under full supervision on the PUMA dataset, benchmarked directly against Akbarpour et al.'s published Dice score
Four semi-supervised methods implemented and compared under matched conditions: Mean Teacher consistency regularization, a semi-supervised GAN, autoencoder pretraining, and confidence-based pseudo-labeling
Simulated label scarcity at 15 and 40 labeled images, with every method evaluated against the same supervised baseline at the same label counts
A domain-diversity ablation sourcing real out-of-domain histopathology data from TCGA-SKCM via the NIH GDC API, to test whether unlabeled-data diversity — not just quantity — affects semi-supervised performance
Status
Supervised baseline (U-Net, full supervision)
Benchmark comparison
Best SSL method at 15 labeled images
Out-of-domain diversity ablation (TCGA-SKCM)
Datasets
- PUMA histopathology dataset — 205 expert-annotated melanoma images (labeled component)
- TCGA-SKCM (The Cancer Genome Atlas, skin cutaneous melanoma), sourced via the NIH GDC API — out-of-domain unlabeled data for the follow-up diversity experiment
Stack
Current focus
- GAN-based semi-supervision reached near-parity with full supervision using only 15 labeled images (0.869 vs. 0.870 Dice) — the strongest of the four methods tested at that label count
- A stalled result in one setting was traced to insufficient domain diversity in the unlabeled pool, not a weakness in the method itself
- Sourcing real out-of-domain histopathology data from TCGA-SKCM to directly test that diagnosis: added diversity partially closed the gap under controlled ablation, but did not surpass supervised training
- The four SSL techniques were evaluated under genuinely matched conditions — same baseline, same label counts, same evaluation protocol — rather than each being reported against its own best-case setup
Stated plainly
- Results are from the PUMA dataset and a TCGA-SKCM out-of-domain ablation — not yet validated on a held-out clinical cohort
- The manuscript describing this work is still being written; a preprint or venue hasn't been finalized yet
- Computer-aided diagnosis here is intended as pathologist decision support, not a validated standalone clinical tool
Earlier work
Other published research
MedInsight: Evaluating Retrieval-Augmented Vision-Language Models for Evidence-Grounded Medical Image Understanding
MedInsight tests whether retrieval-augmented prompting improves a vision-language model's accuracy on medical visual question answering, using only openly licensed, non-credentialed data. A CLIP-embedded FAISS index over ~90,000 ROCOv2 radiology image-caption pairs supplies retrieved evidence captions to a pretrained BLIP-2 model, and that retrieval-augmented condition is compared against the same model answering with no retrieved evidence at all — same checkpoint, same decoding settings, same scoring code, on the official VQA-RAD test split (451 clinician-authored questions), with paired bootstrap significance testing rather than a single reported delta.
AI-Based Depression Detection from Social Media Using LSTM Networks
This research investigates whether Long Short-Term Memory (LSTM) networks can pick up linguistic and behavioral indicators of depression in social media text more reliably than simple bag-of-words baselines. NLP preprocessing (tokenization, cleaning, stopword removal, embedding generation) feeds a sequence model trained to classify posts as depression-indicative or not, with class-imbalance handling and standard classification metrics.