ALLNet: Acute Lymphoblastic Leukemia Detection Using Lightweight Convolutional Networks

Methods for detecting Acute Lymphoblastic (or Lymphocytic) Leukemia (ALL) based on the analysis of blood images are being increasingly researched in the context of Computer Aided Diagnosis (CAD) systems, which help the pathologist in performing the diagnosis. Within CAD systems, approaches using Deep Learning (DL) and Convolutional Neural Networks (CNN) currently exhibit the highest accuracy in detecting the presence of lymphoblasts, which indicate the possible presence of ALL. Recently, approaches based on histopathological transfer learning have been proposed to increase the accuracy of ALL detection in the presence of databases with a small number of samples, by pretraining the CNN on histopathological data instead of using general-purpose datasets such as ImageNet. However, all the approaches in the literature consider CNN architectures with an extremely high number of learnable parameters, which easily tend to overfit. To compensate for these drawbacks, in this paper we propose ALLNet, the first approach in the literature for ALL detection using a lightweight architecture based on fixed binary kernels that replicate the Local Binary Patterns and that uses only ≈ 1.6% of the learnable parameters of a traditional CNN. We evaluated our approach on a public ALL database, achieving better results with respect to the state of the art in terms of classification accuracy.