Malaria infections affect millions of people worldwide, having a considerable expression in lowresource locations, such as African countries. The limitations presented by the currently available diagnostic tools – rapid diagnostic tests (RDTs) and microscopic examination – emphasise the need for a portable and automatic diagnostic method. The MalariaScope system proposed by Rosado et al. integrates algorithms for the automated examination of blood smears with a 3D-printed portable microscope – the µSmartScope – and a smartphone application, providing a promising and practical device, whose performance can be improved in terms of reliability and execution time. In this context, the present dissertation is focused on the application of deep learning approaches for the identification of malaria parasites (MPs) in microscopic thin blood smear images and their characterisation in terms of species and life-cycle stage, to refine the already developed system and attenuate some of the obstacles reported.
In order to accomplish the proposed objectives, pipelines that combined the detection and classification of the parasites in different manners were envisioned and tested. In the selected pipeline, the images acquired first go through pre-processing operations to standardise image size and reduce computational effort. The processed images are used as input to a CNN-based object detection module, which not only detects the parasite objects in the images, but also determines their life-cycle stage (trophozoite, schizont or gametocyte). In the classification module, each detected parasite is used as input to a species classifier specific of each stage, whose output corresponds to its species (P. falciparum, P. vivax, P. ovale or P. malariae). An additional class for artefacts is considered to filter possible incorrectly detected objects. The proposed system does not provide any information regarding the species of schizont objects because the scarcity of schizonts available in the dataset hindered the application of CNN classifiers to this type of parasites.
The detection approach implemented (single-shot multi-box detector with a MobileNet v2 as a feature extractor) presented a high sensitivity to MP objects, with only 5 (1%) undetected parasites. Its localisation performance is comparable to the one of the baseline segmentation approach suggested by Rosado et al., and tests in a mobile platform demonstrated that it allows a fast processing of the thin smear images; yet, the sensitivity to schizonts and gametocytes can still be improved. Therefore, future work should explore alternative model architectures, as well as the inclusion of more examples of the under-represented classes. In case the latter is not possible, additional data augmentation techniques, namely auto-encoders and generative adversarial networks, ought to be investigated.
The CNN models used for the classification module, corresponding to 5-class MobileNet v2 architectures, presented reasonable overall performances (F1 measures of 0.30 and 0.62 for the classification of trophozoites and gametocytes, respectively). Higher evaluation metrics were obtained for artefacts and P. falciparum instances, showing that the algorithm trained was able to segregate objects incorrectly identified in the detection module and discriminate P. falciparum parasites with satisfactory reliability. Despite the results attained, a direct comparison with the baseline classification methodology demonstrated that a more robust classification algorithm is needed, encouraging the exploration of more training and model settings for this module.
Thus, the pipeline proposed for the diagnosis of malaria in thin blood smear images presented promising performances for both the detection and classification tasks, but is not sufficiently mature for its integration in the MalariaScope framework yet. The potential demonstrated by the algorithms analysed motivates a more thorough investigation and the execution of additional tests in terms of inference time and memory requirements, to assess the practical applicability of this novel approach.
Author: Ana Sampaio
Type: MSc thesis
Partner: Faculdade de Engenharia da Universidade do Porto