New AI Model Learns Diagnostic Skill from Pathologists' Eyes

RadAISlice Newsletter

Weekly Updates in Radiology AI

Good morning, there. Researchers trained an AI on pathologists’ eye movements to improve biopsy analysis accuracy.

By leveraging real-time expert gaze data, this AI model achieves high accuracy with less labor-intensive annotation. This approach lowers barriers for building sophisticated medical AI, inspiring similar advances in radiology where annotation is often a bottleneck. Such scalable methods could speed multimodal AI research and clinical deployment across medical image analysis.

Here's what you need to know about Radiology AI last week:

• AI Learns From Expert Eyes to Advance Biopsy Image Analysis

• Habitat AI Model Boosts Lung Nodule Risk Stratification on LDCT

• Hybrid AI-Human Mammography Workflow Cuts Workload Safely

• AI Model Predicts Multiple Genetic Biomarkers from Colorectal Slides

• Plus: 3 FDA approved devices & 4 new papers.

🔍 NEW WEBSITE FEATURE

We've added filters to our papers section! You can now search papers by modality, body part, task type, and more.

Try it now at xrayinterpreter.com/papers →

🧑‍⚕️ AI Learns From Expert Eyes to Advance Biopsy Image Analysis

Image from: EurekAlert

RadAISlice: Researchers trained AI to recognize diagnostic clues by tracking pathologist gaze during slide review.

The details:

• Model trained on 5,881 skin lesion slides with expert eye movements as supervisory signal

• AI exceeded the next-best model’s accuracy by 5.5% on external test data

• Reduces need for labor-intensive manual labeling in digital pathology

• Combining gaze-driven annotation with other models boosts diagnostic accuracy

• Aims to enable future multimodal and personalized AI-driven diagnosis

Key takeaway: Eye-tracking-based annotation offers a scalable path for infusing human diagnostic expertise into AI tools, both in pathology and radiology.

🫁 Habitat AI Model Boosts Lung Nodule Risk Stratification on LDCT

RadAISlice: A 'habitat' AI model refines risk classification of subsolid lung nodules on lung screening CT.

The details:

• Tested multiple models on 747 subjects with 834 lung adenocarcinomas

• Habitat and radiomic models reached AUC of 0.92, outperforming 2D model

• 'Habitat imaging' segments nodules by spatial heterogeneity for better analysis

• Combined model achieved the highest accuracy (AUC 0.93)

• Interpretability helps reduce variability among radiologists

Key takeaway: AI tools quantifying nodule heterogeneity can sharpen lung cancer risk stratification, supporting early, tailored screening and reducing subjective variability.

🎓 Hybrid AI-Human Mammography Workflow Cuts Workload Safely

RadAISlice: Hybrid AI and radiologist interpretation in mammography reduced workload by 38% while preserving accuracy.

The details:

• Study covered over 41,000 mammograms from Dutch screening program

• AI interpreted confident cases; uncertain cases went to radiologists

• Recall and cancer detection rates matched standard double reading

• AI uncertainty estimation was key for triage

• Just one cancer missed by AI but caught by human reviewer

Key takeaway: Thoughtful integration of AI with human expertise can meaningfully reduce radiologist workload in breast cancer screening with no drop in diagnostic standards.

🔬 AI Model Predicts Multiple Genetic Biomarkers from Colorectal Slides

Image from: EurekAlert

RadAISlice: A multi-target AI predicts several genetic colorectal cancer biomarkers from standard histology slides.

The details:

• Analyzed nearly 2,000 slides across seven diverse cohorts in Europe and the US

• AI predicted multiple key mutations simultaneously at or above single-target accuracy

• Can identify shared image patterns tied to several mutations

• Allows faster, scalable, lower-cost prescreening before molecular testing

• Work featured cross-institutional, international collaboration

Key takeaway: Multi-target AI models can streamline genetic biomarker detection, accelerating precision oncology and informing multimodal radiopathology strategies.

🏛️ FDA Clearances

• K250755 - DS Core Diagnosis by Dentsply Sirona is FDA cleared AI software for radiological image analysis.

• K251747 - EOS imaging received clearance for VEA Align; spineEOS, an AI-driven X-ray spine alignment and analysis system.

• K250170 - Brightonix Imaging’s PHAROS, a PET system for 3D body imaging, has gained FDA 510(k) clearance.

📄 Fresh Papers

• doi:10.1101/2025.08.14.25333618 - GigaHeart is a cardiac CT foundation model enabling highly accurate heart size prediction for transplantation.

• doi:10.1088/2516-1091/adfeab - Review shows deep learning is driving advances in every phase of image-guided tumor ablation and patient outcomes.

• doi:10.1016/j.jocmr.2025.101945 - ScarNet foundation model enables automated, accurate segmentation of myocardial scar in cardiac MRI images.

• doi:10.1007/s00330-025-11924-3 - Study finds large language models can assist differential diagnosis of brain tumors from MRI reports, but trail radiologists.

• Browse 227 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Former US Surgeon General Jerome Adams joins Eko Health to champion AI-powered cardiac and pulmonary diagnostics.

• Dutch researchers show hybrid AI can safely reduce mammography workload 38% without impacting cancer detection.

• Estonia’s Better Medicine raised $7M to expand its kidney CT AI platform and prepare for US market entry.

• A machine learning model using chest CT accurately diagnoses and grades COPD, offering new clinical value.

• Brazilian researchers urge that physician stewardship is essential for ethical patient-centered AI in medicine.

• Read 13 fresh radiology AI news articles for last week.

That's it for today!

👋 We'd love your thoughts. Just hit Reply and tell us what you want to see in the next Slice!

See you next week.