Abstract:
Millions of people around the world are afflicted by malaria, a potentially lethalillness that frequently goes undiagnosed. Although microscopy is still frequentlyused for diagnosis, it greatly depends on professional interpretation. This papersuggests a deep learning-based categorization method that makes use of a VisionTransformer (ViT) to improve speed, accuracy, and scalability. We classifiedmalaria- infected blood smear images from two datasets using a ViT-B/16model pretrained on ImageNet. The images were divided into four classes:parasitized, uninfected, slides_with_positive_cells, and slides_with_negative_cells. To ensure an equal distribution of classes, all photos were scaledto 224 x 224 pixels, normalized, and divided into training (70%), validation(15%), and test (15%) groups. Each image is split up into 16x16 patches by theViT architecture, which then processes them using 12 transformer encoderlayers. To generate predictions across the four categories, a softmax layer wasapplied after a bespoke linear head took the place of the original classifier.Training took place over 10 epochs using the AdamW optimizer, cross-entropyloss, and a batch size of 32. Following each epoch, performance was monitoredon the validation set, and all metrics were recorded for later examination. Weintroduced new output neurons to the classification head (for ring forms,trophozoites, etc.) while freezing the backbone to facilitate incremental learningwhile maintaining previously learned characteristics. Because only the new headwas educated, the model was able to adjust without losing its existinginformation. Lastly, classification reports, confusion matrices, and one-vs-allROC curves with AUC scores for each class were used to assess the model'sperformance on the test set
Page(s):
100-100
DOI:
DOI not available
Published:
Journal: 4th International Conference of Sciences “Revamped Scientific Outlook of 21st Century, 2025” , November 12,2025, Volume: 1, Issue: 1, Year: 2025
Keywords:
malaria
,
CNN
,
Incremnetal learning
,
Transformer