Goal: Develop a reliable kidney lesion detector, able to handle sequences of data coming from CT scans. 5 kidney lesions were included for detection. Problem: Long waiting times for the patient before receiving a diagnostic. Time consuming work for the radiologists, manually detecting the kidney lesions and omissions because of high workload.
Problem: Long waiting times for the patient before receiving a diagnostic. Time consuming work for the radiologists, manually detecting the kidney lesions and omissions because of high workload.
Challenges
Machine Learning and Predictive Analytics:
Data imbalance and having significantly less lesions in the underrepresented classes.
Data leakage was a topic of concern given the sequential structure of the data, therefore special attention was required when creating the train – validation – test sets and in cross validation.
The annotations coming from multiple sources (radiologists) had to be manipulated to meet a common ground. For example a nodule cluster could be annotated with one, large bounding box or with individual bounding boxes around each nodule.
Difficulty in detecting lesions on the edge of the organ.
Algorithms
Machine Learning and Predictive Analytics:
The problem was tackled as an object detection problem, therefore optimizing for two outputs: the coordinates of the bounding box and the correctness of the label assigned to a bounding box.
The approach was to use, as a backbone, a very deep neural network and to adapt its pre-trained parameters through fine-tuning such that it accommodates medical images.
For performance evaluation, both bounding box overlap and label matching were assessed.
The metrics under observation were: IoU (Intersection over Union), precision, recall, f1 score and MAP (mean average precision).
This project had a strong hyperparameter tuning component, requiring extensive experimentation with advanced concepts: one cycle learning rate, learning rate schedulers, optimizers, kernel size and activation functions.
Results
The solution was able to detect the lesions in under 30 seconds per patient (on a sequence of 40 CT slices), with a precision and recall of >85%, making it suitable for internal hospital use.
Impact
Improve the patient experience by reducing the diagnostic waiting times.
Streamline the radiologist’s workflow, empower novice radiologists in diagnosing and improve their ramp-up process.
