Training Dataset Pruning Algorithm with Evaluation on Medical Datasets

The machine learning (ML)-community is constantly trying to improve classification models, and one of the common approaches is to enlarge a learning dataset. However, this can be demanding in medical applications, especially in MLassisted colorectal cancer (CRC) detection, where the publicly available data is limited, and researchers usually rely on small internal CRC datasets. In this paper, an iterative algorithm is proposed, which can be used two-fold: first, to follow the trend of the model performance curve by increasing the input data and to test the cost-effectiveness of the data incrementation process, or second, to obtain an efficient training dataset pruning. The input dataset used for algorithm evaluation consisted of three publicly available sets of CRC images: Colorectal Histology, NCT- CRC-HE-100K, and CRC-VAL-HE-7K. The strategy of using merged datasets from different sources was revisited and used as a benchmark for the current study. The accuracy obtained with the MobileNetV2 model was $\sim 97 \%$. As concluded from the results with the implemented algorithm, the classification efficiency remained high (drop was $\sim 5 \%$), while a large percentage of data (87.50% out of 76968) was pruned. However, this paper offers interesting preliminary research and a valuable starting point for further investigation. The focus was on the pruning algorithm implementation instead of the specific medical application. It is shown that the algorithm is practical and general enough to be applied to any training dataset. Also, it enables one to examine the trade-off between the input data reduction and the baseline classification performance.