ACCEPTED PAPER · ICADCML 2026
Reading knee texture for osteoporosis, not just shape
With Arup Kumar Pal, Manish Raj, and Jitesh Pradhan, I worked on texture-based feature extraction paired with a CBAM-enhanced U-Net for automated knee osteoporosis detection. The model supports both a binary reading and a three-class severity split.
- Encoder
- ResNet + CBAM
- Context module
- ASPP
- Classification
- Binary and 3-class
- Input size
- 128 × 128
THE QUESTION
Bone texture carries signal that shape alone misses
Osteoporosis shows up as a loss of bone density and a change in the internal trabecular pattern, not only as a change in outline. A model that only looks at overall shape can miss that texture, so we wanted an approach that extracts texture descriptors explicitly and lets a learned encoder build on top of them.
The classification target is set up two ways in the codebase: a binary read for presence or absence, and a three-class read for finer severity grading.
METHOD
Texture features feeding an attention U-Net
- 01
Texture feature extraction
Hand-crafted texture descriptors are extracted from each knee image, then reduced with PCA: 75 components for the binary setting, 90 for the three-class setting.
- 02
ResNet encoder with CBAM
A ResNet-based encoder builds learned image features. A Convolutional Block Attention Module is applied so the network weights the channels and regions that matter most for bone texture.
- 03
ASPP for multi-scale context
An Atrous Spatial Pyramid Pooling module gathers context at several receptive field sizes before the features reach the decoder.
- 04
Decoder and classification
The decoder combines the multi-scale features with the reduced texture descriptors, and a classification head outputs the binary or three-class osteoporosis label.
TRAINING SETUP
Default configuration
CURRENT EVIDENCE
Accepted, with a benchmark table still to come
No public accuracy table is published on this page yet
The paper has been accepted at ICADCML 2026, and the repository contains the full texture extraction, model, and training code. A public leaderboard-style results table has not been added here, and one will follow once the camera-ready results are finalized.
The seed is fixed in the default configuration, which keeps individual training runs reproducible, but no claim of clinical validation is made at this stage.
REPRODUCE
Code and paper status
Texture-based Feature Extraction and CBAM-Enhanced U-Net for Automated Knee Osteoporosis Detection · Roy, Pal, Raj, Pradhan · accepted at ICADCML 2026.