Development of a chest X-ray machine learning convolutional neural network model on a budget and using artificial intelligence explainability techniques to analyze patterns of machine learning inference
Bibliographic record
Abstract
Objective: Machine learning (ML) will have a large impact on medicine and accessibility is important. This study's model was used to explore various concepts including how varying features of a model impacted behavior. Materials and Methods: This study built an ML model that classified chest X-rays as normal or abnormal by using ResNet50 as a base with transfer learning. A contrast enhancement mechanism was implemented to improve performance. After training with a dataset of publicly available chest radiographs, performance metrics were determined with a test set. The ResNet50 base was substituted with deeper architectures (ResNet101/152) and visualization methods used to help determine patterns of inference. Results: Performance metrics were an accuracy of 79%, recall 69%, precision 96%, and area under the curve of 0.9023. Accuracy improved to 82% and recall to 74% with contrast enhancement. When visualization methods were applied and the ratio of pixels used for inference measured, deeper architectures resulted in the model using larger portions of the image for inference as compared to ResNet50. Discussion: The model performed on par with many existing models despite consumer-grade hardware and smaller datasets. Individual models vary thus a single model's explainability may not be generalizable. Therefore, this study varied architecture and studied patterns of inference. With deeper ResNet architectures, the machine used larger portions of the image to make decisions. Conclusion: An example using a custom model showed that AI (Artificial Intelligence) can be accessible on consumer-grade hardware, and it also demonstrated an example of studying themes of ML explainability by varying ResNet architectures.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".