Related Experiment Video
Updated: Oct 21, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Multitask Deep Learning for Segmentation and Classification of Primary Bone Tumors on Radiographs
Claudio E von Schacky1, Nikolas J Wilhelm1, Valerie S Schäfer1
1From the Department of Radiology (C.E.v.S., V.S.S., Y.L., F.G.G., S.C.F., F.T.G., M.R.M., K.W., A.S.G.), Department for Orthopedics and Orthopedic Sports Medicine (N.J.W., C.K., R.v.E., R.B.), and Institute of Pathology (C.M.), Klinikum Rechts der Isar, Technische Universität München, Ismaninger Str 22, 81675 Munich, Germany; and the Department of Diagnostic and Interventional Radiology, Medical Center-University of Freiburg, Faculty of Medicine, Freiburg, Germany (M.J., P.M.J., M.F.R.).
A new deep learning (DL) model accurately identifies primary bone tumors on radiographs, aiding diagnosis. This artificial intelligence tool simultaneously locates, segments, and classifies bone tumors, improving workflow efficiency.
Area of Science:
- Artificial Intelligence in Radiology
- Medical Imaging Analysis
- Oncology Diagnostics
Background:
- Primary bone tumors require accurate assessment for effective treatment.
- Radiographs are a common imaging modality for bone tumors.
- An artificial intelligence (AI) model can potentially enhance diagnostic workflows.
Purpose of the Study:
- To develop a multitask deep learning (DL) model.
- The model aims for simultaneous bounding box placement, segmentation, and classification of primary bone tumors.
- Evaluate the model's performance on radiographs.
Main Methods:
- Retrospective analysis of bone tumors on radiographs from 2000-2020.
- Histopathologic findings served as the reference standard for diagnosis.
- Split-sample validation (70% training, 15% validation, 15% test) and external geographic validation were employed.
Main Results:
- The multitask DL model achieved 80.2% accuracy, 62.9% sensitivity, and 88.2% specificity for tumor classification.
- The model demonstrated an Intersection over Union (IoU) of 0.52 for bounding box placement and a Dice score of 0.60 for segmentation.
- Model performance in classification was superior to radiology residents and comparable to fellowship-trained radiologists.
Conclusions:
- The developed multitask DL model enables accurate, simultaneous bounding box placement, segmentation, and classification of primary bone tumors.
- This AI tool shows promise for assisting in the diagnostic workflow of bone tumors.
- The model's performance suggests its potential utility in clinical radiology settings.
