Sasha-Gay Wright
Research Subjects
Sasha-Gay Wright has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Abu Salam
Universitas Dian Nuswantoro
2 shared publications - Antonius Gunawan Santoso
Department of Radiology, Faculty of Medicine , Diponegoro University , Semarang , Indonesia
2 shared publications - Cinantya Paramita
Universitas Dian Nuswantoro
2 shared publications - Erwin Yudi Hidayat
Universitas Dian Nuswantoro
2 shared publications - Moch Arief Soeleman
Universitas Dian Nuswantoro
2 shared publications - Pulung Nurtantio Andono
Universitas Dian Nuswantoro
2 shared publications
Latest Publications
- Open Access
- 18 June 2026
- Published by Elsevier in SSRN Electronic Journal
Investigating Cross-Hospital Generalization of Chest Radiograph Lesion Detection Through Public-to-Institutional Source-Holdout Evaluation
Robust cross-hospital generalization remains a major challenge in chest radiograph lesion detection, especially when detectors trained on mixed public and institutional sources are transferred to an unseen hospital. This study investigated whether a Digital Imaging and Communications in Medicine (DICOM)-aware adaptive harmonization branch, Adaptive HyFusion, improves public-to-institutional source-holdout transfer compared with standard preprocessing. A fixed Faster R-CNN detector with a ResNet-50-FPN backbone was evaluated under a harmonized five-class lesion ontology using VinDr-CXR, MS-CXR, and two non-public Indonesian hospital cohorts. Scenario 2A, using RSUD Dr. Adhyatma, MPH Jawa Tengah for institutional development and RSUD Sunan Kalijaga Kabupaten Demak as the locked held-out test source, was the main analysis; Scenario 2B reversed the hospital roles as a sensitivity analysis. Performance was assessed using held-out mean average precision at an intersection-over-union threshold of 0.50 (mAP50), class-wise AP50, recall, precision, false positives per image, and validation-to-held-out degradation. In Scenario 2A, Adaptive HyFusion achieved comparable held-out mAP50 to baseline preprocessing (0.1169 ± 0.0460 vs. 0.1156 ± 0.0545), with higher recall (0.1282 ± 0.0513 vs. 0.1026 ± 0.0128), higher precision (0.0924 ± 0.0053 vs. 0.0728 ± 0.0143), and lower false positives per image (1.2215 ± 0.3754 vs. 1.2673 ± 0.1866). In Scenario 2B, Adaptive HyFusion achieved higher held-out mAP50 (0.1236 ± 0.0377 vs. 0.1070 ± 0.0102), higher precision, lower false-positive burden, and smaller validation-to-held-out mAP50 degradation, although recall was slightly lower. Class-wise effects remained heterogeneous. These findings support DICOM-aware adaptive preprocessing as a meaningful component of cross-hospital generalization strategy rather than a universally superior detection solution.
- 18 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessInvestigating Cross-Hospital Generalization of Chest Radiograph Lesion Detection Through Public-to-Institutional Source-Holdout Evaluation
Robust cross-hospital generalization remains a major challenge in chest radiograph lesion detection, especially when detectors trained on mixed public and institutional sources are transferred to an unseen hospital. This study investigated whether a Digital Imaging and Communications in Medicine (DICOM)-aware adaptive harmonization branch, Adaptive HyFusion, improves public-to-institutional source-holdout transfer compared with standard preprocessing. A fixed Faster R-CNN detector with a ResNet-50-FPN backbone was evaluated under a harmonized five-class lesion ontology using VinDr-CXR, MS-CXR, and two non-public Indonesian hospital cohorts. Scenario 2A, using RSUD Dr. Adhyatma, MPH Jawa Tengah for institutional development and RSUD Sunan Kalijaga Kabupaten Demak as the locked held-out test source, was the main analysis; Scenario 2B reversed the hospital roles as a sensitivity analysis. Performance was assessed using held-out mean average precision at an intersection-over-union threshold of 0.50 (mAP50), class-wise AP50, recall, precision, false positives per image, and validation-to-held-out degradation. In Scenario 2A, Adaptive HyFusion achieved comparable held-out mAP50 to baseline preprocessing (0.1169 ± 0.0460 vs. 0.1156 ± 0.0545), with higher recall (0.1282 ± 0.0513 vs. 0.1026 ± 0.0128), higher precision (0.0924 ± 0.0053 vs. 0.0728 ± 0.0143), and lower false positives per image (1.2215 ± 0.3754 vs. 1.2673 ± 0.1866). In Scenario 2B, Adaptive HyFusion achieved higher held-out mAP50 (0.1236 ± 0.0377 vs. 0.1070 ± 0.0102), higher precision, lower false-positive burden, and smaller validation-to-held-out mAP50 degradation, although recall was slightly lower. Class-wise effects remained heterogeneous. These findings support DICOM-aware adaptive preprocessing as a meaningful component of cross-hospital generalization strategy rather than a universally superior detection solution.