Automated defect detection in ultrasound probes using patch-based machine learning algorithm

Quality control is considered one of the most important tasks by manufacturing industries, as being able to guarantee certain quality levels is necessary in order to meet customer expectations, reduce waste and costs. Requiring an inspection for each of the products manufactured, being able to have a tool capable of automating the detection of possible defects would simplify this procedure, making it faster and more optimised. This is the specific case of the company Esaote, which carries out accurate quality control for each of its ultrasound probes. In particular, concerning the inspection of the shell, that is currently carried out manually by the staff. To create a tool capable of automating this inspection, a defect detection algorithm based on machine learning is proposed. The algorithm processes the images according to a patch-based strategy, following three steps: splits the image in patches, classifies and localizes the defects in the patches, and finally combines the results to obtain the complete list of all the defects present. Performance was analysed for different configurations related to the three steps, with particular emphasis on the percentage of overlap between patches, the total number of patches to be analysed, and the voting algorithm used to combine results between several patches. The best configuration identified an accuracy of 83.63%, an f-score of 89.87%, a precision of 81.60%, a recall of 88.97%, an AP of 77.63% and finally an AUC of 83.48%.