From Classification to Localization and Clinical Validation: Large-Scale Development of a Deep Learning System for Thoracic Disease Detection on Chest Radiographs in Thailand

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Advanced, extended

Summary

Inspectra CXR version 5 is a deep learning system developed for multi-label thoracic disease classification and weakly supervised lesion localization on chest radiographs in Thailand. Built on a DenseNet-121 backbone augmented with Attend-and-Compare Modules (ACM) and a Probabilistic Class Activation Map (PCAM) aggregation layer, the model was trained on 874,858 frontal chest radiographs from Siriraj Hospital. On a held-out in-domain test set of 19,871 cases, it achieved a mean AUROC of 0.994, with 92.4% sensitivity and 98.6% specificity across nine conditions. The system demonstrated robust cross-site generalization, maintaining a mean AUROC of 0.970 on 5,992 cases from 13 diverse Thai hospitals. For localization, it attained a mean lesion-localization fraction (LLF) of 77.9% on 4,549 radiologist-annotated cases. Clinical usability evaluation with five thoracic radiologists showed 93.6% classification concordance, 94.7% localization concordance, and an "excellent" mean System Usability Scale (SUS) score of 89.

Key takeaway

For Machine Learning Engineers developing medical imaging AI, this study demonstrates that focusing on local data adaptation and comprehensive validation is crucial. You should prioritize multi-task architectures that provide both classification and weakly supervised localization, as these enhance clinical utility and radiologist trust. Ensure your validation includes diverse real-world sites and formal usability assessments like the SUS to prove practical applicability and acceptance.

Key insights

A locally adapted deep learning system for CXR achieves high accuracy, cross-site generalization, and radiologist trust through simultaneous classification and localization.

Principles

• Local adaptation significantly improves model accuracy for specific populations.
• Simultaneous classification and localization enhances clinical interpretability.
• Comprehensive validation includes in-domain, cross-site, and usability studies.

Method

The system uses a DenseNet-121 + ACM encoder, external attention decoder, and PCAM aggregation layer for multi-task classification and weakly supervised localization.

In practice

• Deploy localization-capable CXR AI as an assistive second reader.
• Use multi-site validation to ensure model robustness across diverse clinical settings.
• Incorporate System Usability Scale (SUS) for assessing clinical acceptance.

Topics

• Deep Learning
• Chest Radiography
• Lesion Localization
• Clinical Validation
• System Usability Scale
• Thoracic Disease Detection

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Scalable Training of Spatially Grounded 2D Vision-Language Models for Radiology — Spatially grounded radiology VLMs can be trained at scale without manual annotations using LLM-based curation.
• Bisakah AI Mendeteksi COVID-19 dari Chest X-Ray? — CNNs can effectively classify COVID-19 from chest X-rays, offering a rapid diagnostic aid.
• Bowel Obstruction Detection and Localization on Abdominal CT with Deep Learning — A multi-task deep learning framework detects bowel obstruction and localizes its transition zone on CTs, enhancing diagnostic workflow.
• Diffusion Attention Expert Model for Predicting and Semi-automatic Localizing STAS in Lung Cancer Histopathological Images — DAEM accurately detects STAS in lung cancer histopathology, offering strong generalizability and TME-based localization.
• When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT — Input dimensionality significantly impacts model performance and stability in volumetric medical imaging.
• Towards Global AI-Driven Cervical Cancer Screening — A multi-task deep learning model validated on diverse global data outperforms experts in cervical cancer screening.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.