Time-series Medical Image Classification
Image Classification Model for Temporal Disease Progression of Chest X-ray dataset
Introduction
Disease progression modeling (DPM) uses mathematical and scientific principles to describe the quantitative progression of a disease over time. One such task is to predict the three states of disease progression (improving, stable, or worsening) based on current and past chest X-ray images. This project fine-tunes and evaluates the pre-trained Torch X-ray Vision model for this temporal image classification task. The code is available on GitHub.
Experiments and Results
The following tables summarize the performance of different feature extraction methods and classification models for the five diseases in the dataset:
Table 1: Performance of Models on Disease Progression Task
| Model/Feature Extractor | Edema | Consolidation | Pleural Effusion | Pneumothorax | Pneumonia | Average Accuracy |
|---|---|---|---|---|---|---|
| DenseNet-RSNA | 45.28% | 45.0% | 60.98% | 50.0% | 61.70% | 53.70% |
| DenseNet-MIMIC-CH | 45.28% | 42.5% | 57.31% | 54.76% | 63.83% | 53.40% |
| DenseNet-CheX | 47.17% | 50.0% | 62.20% | 54.76% | 63.83% | 56.40% |
| Logistic Regression | 44.08% | 51.00% | 39.02% | 37.21% | 56.13% | 45.49% |
Table 2: Prediction on Flipped Input (Logistic Regression Model)
| Original Prediction | Stable | Improving | Worsening | Prediction on Flipped Input |
|---|---|---|---|---|
| Stable | 54 | 45 | 0 | Stable |
| Improving | 21 | 92 | 7 | Improving |
| Worsening | 4 | 32 | 3 | Worsening |
Table 3: Prediction on Flipped Input (Model 2)
| Original Prediction | Stable | Improving | Worsening | Prediction on Flipped Input |
|---|---|---|---|---|
| Stable | 96 | 33 | 1 | Stable |
| Improving | 39 | 58 | 22 | Improving |
| Worsening | 1 | 13 | 1 | Worsening |
Table 4: Prediction on Flipped Input (Model 3)
| Original Prediction | Stable | Improving | Worsening | Prediction on Flipped Input |
|---|---|---|---|---|
| Stable | 48 | 18 | 10 | Stable |
| Improving | 12 | 41 | 29 | Improving |
| Worsening | 9 | 21 | 76 | Worsening |
Future Work
To improve the model, future research can explore semi-supervised learning techniques to handle missing data more effectively and investigate the use of lateral chest X-rays for enhanced predictive capabilities. Hyperparameter tuning of the Vision Transformer model can also lead to performance gains.
For more details, check the GitHub Repository.