Adaptive Multimodal Fusion of Histopathology and Radiographs for Osteosarcoma Classification Using EfficientNet and Gated Decision Integration
IEEE Access, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3716879
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Anahtar Kelimeler: Adaptive gating network, Decision-level fusion, Histopathology, Multimodal deep learning, Osteosarcoma classification
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
Osteosarcoma diagnosis relies on complementary information from radiographic assessment of bone-level structural abnormalities and histopathological assessment of cellular tissue morphology. In public research datasets, however, these two modalities are usually acquired independently and cannot be assumed to have patient-level correspondence. This study proposes an adaptive decision-level fusion framework for three-class osteosarcoma classification, including non-tumor, viable tumor, and necrotic tumor, using unpaired histopathology tiles and radiographic images. Two EfficientNet-B0 networks were trained independently: one on histopathological data from the TCIA Osteosarcoma Tumor Assessment dataset, containing 1,144 tiles from four patients, and one on radiographic data from the Kaggle Bone Tumor Classification dataset, containing 180 images without patient-level identifiers. A shallow gating network learned per-sample fusion weights from the concatenated softmax probability vectors of both branches. Leave-one-patient-out cross-validation across all four TCIA patients yielded a mean accuracy of 95.89%, mean macro F1 of 0.95, and mean adaptive weight α̂ of 0.38. On the primary held-out test patient P004, containing 171 tiles, the adaptive model achieved 97.08% accuracy, macro F1 = 0.97, and AUROC = 0.99, compared with 94.15% for fixed-weight fusion, α = 0.25, and 94.15% for the max-confidence selection baseline. The McNemar test did not reach statistical significance, p = 0.074; therefore, this result should be interpreted as a promising observed accuracy trend rather than statistically confirmed superiority. To assess histopathology-branch robustness under institutional domain shift, the histopathology branch was evaluated with a uniform radiograph prior on an independent cohort from Hitit University Faculty of Medicine, including 312 patients and 1,248 histopathology images, with Ethics Approval No. 254, dated 17 June 2020. This evaluation yielded 96.23% binary tumor-detection accuracy, 95% CI: [94.80%, 97.50%], macro F1 = 0.96, AUROC = 0.99, and ECE = 0.038. Because this cohort did not include paired radiographs, these results evaluate histopathology-branch robustness rather than full two-modality fusion generalization. All results are presented as a methodological proof of concept; larger prospective multi-institutional studies with paired records are required before clinical translation can be considered.