Small Training Datasets for Deep Learning Based Medical Diagnosis
David Härer, Konrad Kraft, Achim Ibenthal · Communications in computer and information science · 2023
In a computer-aided diagnosis framework binary, patch-based classification of malaria from thick blood smear images is investigated. Following a state-of-the-art study on malaria, deep learning and transfer learning, classical networks like ResNet are being adapted and optimized for training by the public dataset of the Automated Laboratory Diagnostics project of John A. Quinn et al. at Makerere University, Uganda. The dataset comprises 2,703 thick blood smear images treated with Fields stain. Furthermore models are cross-validated on other thick and thin blood smear datasets treated with Giemsa stain. Data augmentation with zoomed-in versions of the image tiles in the training dataset and the use of shallow models with relatively few network parameters is evaluated as a solution to the problem of small training databases. Using grid search, optimum hyperparameters for learning rate scheduling are found empirically for each network architecture. The self-developed deep learning models achieve a TP ( T rue- P ositive) rate of 98% at a FP ( F alse- P ositive) rate of 0.9% on the thick blood smear images and a TP rate of 90% at a FP rate of 8% on the thin blood smear images. The maximum validation accuracy over the set of models and hyperparameters is 99.3%.