Aims and objectives
Achieving high levels of accuracy,
needed for clinical relevance,
machine learning algorithms must be have a high degree of model complexity.
In Deep Learning,
the most performant of these algorithms,
these models frequently have millions of free parameters which are individually,
very difficult to interpret.
The inability to understand how a decision is being made presents a major barrier for adoption in the clinic. As a first study to test interpretable machine learning in a radiological...
Methods and materials
For the study we used Thorax CT images from 15 different patients manually labeled by experts as being pathological or normal .
Morphological and texture information is extracted from the image and a model is trained to classify the vertebra based on the metrics.
For training the data were divided into two groups: training and validation.
The models were trained with the training data and only exposed to the validation data once the training was completed.
Since different models can...
The 4 different models each giving performance between 60-80% accuracy on the validation data.
The feasibility scores and validation data accuracy had a strong correlation of 0.75 with an R^2 of 0.9.
Each of the models highlighted different features (shown in figure 2) which demonstrate which features were relevant for making the classification.
Interpretable explanations provide a powerful insight into which criteria are being used to make a complicated medical decision.
Ultimately this approach can enable the selection of better more generalizable models incorporating physician expertise.
The true value of these approaches comes when dealing with more complicated problems and corresponding models.
A Framework for Automated Spine and Vertebrae Interpolation-Based Detection and Model-Based Segmentation.
IEEE Trans Med Imaging.
“Why Should I Trust You?” Explaining the Predictions of Any Classifier.