Imon Banerjee

📘 New @NCICancerStats chapter: We developed a domain-specific #LLM for prostate cancer using 1.8M clinical notes from 15K+ patients. It outperforms generic LLMs—showing the power of tailored training + domain vocab. #AI #ProstateCancer #v=onepage&q&f=false" target="_blank" rel="nofollow noopener">books.google.com/books?hl=en&lr…

#AIinHealthcare #MedicalImaging #KidneySegmentation #DomainAdaptation

New Research Alert: Robust Kidney Segmentation Across Contrast Phases link.springer.com/article/10.100…

📂 Open-source code available: github.com/ramon349/domai…. Let’s build AI tools that generalize better—because patients deserve consistent care, regardless of imaging conditions.

🔍 Why it matters: ✅ Handles contrast-to-non-contrast and arterial-to-venous shifts ✅ Works on both normal and abnormal kidneys ✅ Open-source and data-efficient ✅ Clinically relevant for automated, contrast-invariant kidney health assessment

Trained on multi-phase public datasets and tested on diverse external datasets (including KiTS21, STU, and Mayo Clinic), our model achieved a DICE score of 0.8892, outperforming state-of-the-art baselines like #TotalSegmentator.

Accurate kidney segmentation from #CT is critical—but current models struggle when faced with domain shifts like contrast phase variation or kidney abnormalities. Our work introduces a #domainadaptation approach using a latent space discriminator to overcome these challenges.

@woojinrad Completely agree

I’m Losing All Trust in the AI Industry The AI industry often makes grand promises about rapid progress that its own experts don’t honestly believe. Many companies prioritize profit over genuine innovation, frequently focusing on creating addictive products rather than developing truly helpful ones. Their public messages are confusing and appear to be designed to attract investment rather than build trust. bit.ly/44C1xmk

.@ImonBanerjee6's research addresses fairness in AI, multimodal deep learning, and AI translation for digital healthcare. She'll discuss her work at the July 14 ML4MI seminar, 10am CT on Zoom. @uwsmph ml4mi.wisc.edu/learning-oppor…

Implication: Opportunistic, scalable, and cost-effective screening—especially useful in resource-limited settings.

MACE remains the leading global cause of death. Our MICCAI 2025 work, MOSCARD, fuses CXR+ECG via multimodal causal reasoning for bias-aware risk prediction. Outperforms SOTA on ED & MIMIC (AUC: 0.75–0.83). #MICCAI2025 #AIHealth arxiv.org/abs/2506.19174

Results: Strong generalization and superior AUC vs single modality and foundation models.

Key innovations: Cross-modal co-attention for CXR-ECG alignment Causal inference to handle confounders Dual back-propagation for de-biasing

We developed MOSCARD: a novel multimodal framework aligning CXR with ECG via causal reasoning

Existing risk models often rely on single modalities or clinical scores—limited by bias and incomplete data.

Codebase github.com/OrchidPi/MOSCA…

🔓 The full pipeline and model are publicly available under an academic license—ready to support clinical and research applications. (github.com/dasavisha/Crit…).

📢 New paper: We propose a framework that leverages weak supervision to train LMs for structured information extraction—minimizing annotation cost while maintaining clinical relevance. #NLP #RadiologyAI #WeakSupervision #ClinicalNLP #LLMs nature.com/articles/s4174…

🧠 Evaluation used both: Manual expert annotations LLM-based metrics: G-eval and Prometheus

The fine-tuned model was evaluated on: Internal test set (Mayo Clinic, n=80) External test set (MIMIC-III, n=123) Large-scale validation (MIMIC-IV, n=5000)

🧪 Our two-phase approach fine-tunes LLMs on 15,000 unlabeled Mayo Clinic reports, using heuristics and domain knowledge to guide learning.