A compact and interpretable multi-source framework for heterogeneous medical image classification.

1. # ML-ConvNet: A Compact and Interpretable Multi-Source Framework for Heterogeneous Medical Image Classification
3. [](https://www.python.org/)
4. [](https://pytorch.org/)
5. [](LICENSE)
7. Official repository for the paper:
9. > **A Compact and Interpretable Multi-Source Framework for Heterogeneous Medical Image Classification**
10. > Williams Ayivi, Xiaoling Zhang, Wisdom Xornam Ativi, Francis Sam
11. > University of Electronic Science and Technology of China | Vienna University of Economics and Business
12. > *Scientific Reports* (under review)
14. ---
16. ## Overview
18. ML-ConvNet is an ultra-lightweight convolutional architecture (~4.2K parameters) designed for
19. heterogeneous medical image classification across independent multi-source imaging cohorts.
20. The framework operates without requiring paired or co-registered multi-modal data, making it
21. suitable for realistic fragmented clinical data ecosystems.
23. ### Key components
25. | Component | Description | Parameters |
26. |-----------|-------------|------------|
27. | MBRConv | Multi-Branch Re-parameterized Convolution | ~912 per block |
28. | IWO | Incremental Weight Optimization strategy | 0 (training only) |
29. | FST | Feature Self-Transformation module | ~72 |
30. | HDPA | Hierarchical Dual-Path Attention | ~60 |
31. | LVW | Local Variance Weighted loss | 0 (training only) |
33. **Total inference parameters:** 4,230 (for K=4 classes)
34. **FLOPs:** 924M at 512x512 input resolution
35. **Inference latency:** 2.3 ms (CPU) | 0.21 ms (GPU) | 6.1 ms (Snapdragon 888)
37. ---
39. ## Datasets
41. Three publicly available benchmark datasets are used. Each is evaluated independently.
43. | Modality | Dataset | Classes | Samples | Imbalance |
44. |----------|---------|---------|---------|-----------|
45. | Brain MRI | [Nickparvar Brain Tumor MRI](https://doi.org/10.34740/kaggle/dsv/14832123) | 4 (Glioma, Meningioma, Pituitary, Normal) | 7,023 | Balanced |
46. | Lung CT | [IQ-OTH/NCCD](https://doi.org/10.34740/kaggle/ds/672399) | 3 (Malignant, Benign, Normal) | 1,097 | High (benign minority) |
47. | Chest X-ray | [Guangzhou Pediatric Pneumonia](https://data.mendeley.com/datasets/rscbjbr9sj/2) | 2 (Pneumonia, Normal) | 5,856 | Moderate |
49. Download each dataset and place under `data/` as follows:
51. data/
52. ├── mri/
53. │ ├── Training/
54. │ │ ├── glioma/
55. │ │ ├── meningioma/
56. │ │ ├── notumor/
57. │ │ └── pituitary/
58. │ └── Testing/
59. ├── ct/
60. │ ├── Malignant cases/
61. │ ├── Benign cases/
62. │ └── Normal cases/
63. └── xray/
64. ├── train/
65. │ ├── PNEUMONIA/
66. │ └── NORMAL/
67. └── test/
70. ---
72. ## Installation
73. ```bash
74. git clone https://github.com/W-ayivi/ML-ConvNet.git
75. cd ML-ConvNet
76. pip install -r requirements.txt
77. ```
79. ---
81. ## Repository Structure
83. ML-ConvNet/
84. ├── models/
85. │ ├── mbrconv.py # Multi-Branch Re-parameterized Convolution + IWO
86. │ ├── fst.py # Feature Self-Transformation module
87. │ ├── hdpa.py # Hierarchical Dual-Path Attention
88. │ └── mlconvnet.py # Full network definition (skeleton)
89. ├── losses/
90. │ └── lvw_loss.py # Local Variance Weighted loss
91. ├── utils/
92. │ ├── reparameterize.py # Structural re-parameterization utility
93. │ └── preprocessing.py # Modality-specific preprocessing pipelines
94. ├── evaluation/
95. │ ├── metrics.py # Classification metrics
96. │ └── intervals.py # CI, PI, TI computation
97. ├── data/
98. │ └── dataset.py # Dataset loaders
99. ├── train.py # Training script
100. ├── evaluate.py # Evaluation script
101. ├── requirements.txt
102. └── README.md
105. ---
107. ## Preprocessing
109. All inputs are standardized to 512x512x3. Single-channel acquisitions (MRI, CT)
110. are replicated across three channels. No modality identity label is passed as a
111. conditioning signal at any stage.
113. | Modality | Pipeline |
114. |----------|----------|
115. | MRI | N4ITK bias correction, skull stripping, in-mask z-score, 3-channel replication |
116. | CT | HU windowing [-1000, 400], scale to [0,1], lesion-centric sampling |
117. | X-ray | CLAHE, median filtering, global z-score (fallback min-max), pad to square |
119. On-the-fly augmentation during training: elastic warping, random rotation, Gaussian noise.
120. Minibatches are constructed via round-robin sampling across cohorts to maintain
121. approximate modality balance. Missing modality inputs are replaced by zero tensors.
123. ---
125. ## Training
126. ```bash
127. python train.py --modality mri --data_dir data/mri --epochs 100 --batch_size 32 --lr 1e-4
128. python train.py --modality ct --data_dir data/ct --epochs 100 --batch_size 32 --lr 1e-4
129. python train.py --modality xray --data_dir data/xray --epochs 100 --batch_size 32 --lr 1e-4
130. ```
132. ### Training configuration
134. | Parameter | Value |
135. |-----------|-------|
136. | Optimizer | Adam |
137. | Learning rate | 1e-4 |
138. | Batch size | 32 |
139. | Epochs | 100 |
140. | Weight initialization | Xavier uniform |
141. | Validation | 10-fold stratified cross-validation (patient-level) |
142. | Framework | PyTorch v2.x, Python 3.12.3 |
143. | GPU | NVIDIA RTX 4060 (8GB VRAM) |
145. Structural re-parameterization is applied after training completion within each fold.
146. Numerical equivalence between training-phase and inference-phase outputs is verified
147. with maximum absolute deviation below 1e-6.
149. ---
151. ## Evaluation
152. ```bash
153. python evaluate.py --modality mri --checkpoint checkpoints/mri_fold1.pth --data_dir data/mri
154. ```
156. Metrics reported per fold: Accuracy, Precision, Recall, F1-score, MCC, Cohen's kappa, AUC.
157. Aggregate statistics: mean +/- SD, 95% CI, 95% PI, 95/95 TI across 10 folds.
159. ---
161. ## Results
163. ### MRI Brain Tumor
165. | Metric | Mean +/- SD | 95% CI |
166. |--------|-------------|--------|
167. | Accuracy | 0.982 +/- 0.002 | [0.981, 0.983] |
168. | Macro F1 | 0.980 +/- 0.002 | [0.979, 0.981] |
169. | MCC | 0.973 +/- 0.002 | [0.972, 0.974] |
170. | AUC | 0.996 +/- 0.001 | [0.995, 0.997] |
172. ### CT Lung Nodule
174. | Metric | Mean +/- SD | 95% CI |
175. |--------|-------------|--------|
176. | Accuracy | 0.949 +/- 0.003 | [0.947, 0.951] |
177. | Macro F1 | 0.933 +/- 0.003 | [0.931, 0.935] |
178. | MCC | 0.918 +/- 0.003 | [0.916, 0.920] |
179. | AUC | 0.972 +/- 0.002 | [0.971, 0.973] |
181. ### Chest X-ray
183. | Metric | Mean +/- SD | 95% CI |
184. |--------|-------------|--------|
185. | Accuracy | 0.958 +/- 0.002 | [0.957, 0.959] |
186. | Macro F1 | 0.950 +/- 0.002 | [0.949, 0.951] |
187. | MCC | 0.940 +/- 0.003 | [0.938, 0.942] |
188. | AUC | 0.978 +/- 0.002 | [0.977, 0.979] |
190. ---
192. ## Citation
193. ```bibtex
194. @article{ayivi2025mlconvnet,
195. title={A Compact and Interpretable Multi-Source Framework for Heterogeneous Medical Image Classification},
196. author={Ayivi, Williams and Zhang, Xiaoling and Ativi, Wisdom Xornam and Sam, Francis},
197. journal={Scientific Reports},
198. year={2025},
199. note={Under review}
200. }
201. ```
203. ---
205. ## Acknowledgements
207. The core architectural design of ML-ConvNet is inspired by the MobileIE framework
208. (Yan et al., ICCV 2025). We thank the authors for making their work publicly available.
210. ---
212. ## License
214. MIT License. See LICENSE for details.
216. ---
218. ## Contact
220. Williams Ayivi — [email hidden]

README.md at commit c7c3a6c, under MIT · at the source