Extracting and Classifying Salient Fields of View from Microscopy Slides of Tuberculosis Bacteria

Abstract

Tuberculosis is one of the most serious infectious diseases, and its treatment is highly dependent on early detection. Microscopybased analysis of sputum images for bacilli identification is a common technique used for both diagnosis and treatment monitoring. However, it a challenging process since sputum analysis requires time and highly trained experts to avoid potentially fatal mistakes. Capturing fields of view (FOVs) from high resolution whole slide images is a laborious procedure, since they are manually localized and then examined to determine the presence of bacteria. In the present paper we propose a method that automates the process, thus greatly reducing the amount of human labour. In particular, we (i) describe an image processing based method for the extraction of a FOV representation which emphasises salient, bacterial content, while suppressing confounding visual information , and (ii) introduce a novel deep learning based architecture which learns from coarsely labelled FOV images and the corresponding binary masks, and then classifies novel FOV images as salient (bacteria containing) or not. Using a real-world data corpus, the proposed method is shown to outperform 12 state of the art methods in the literature, achieving (i) an approximately 10% lower overall error rate than the next best model and (ii) perfect sensitivity (7% higher than the next best model).

Automated Methods for Tuberculosis Detection/Diagnosis: A Literature Review

“…Methods that incorporated a preliminary segmentation stage or a hybrid approach, commonly by leveraging CNNs as feature extraction mechanisms and subsequently feeding these feature maps into another classification/regression algorithm such as an SVM, consistently attained superior results [57,60]. In addition, akin to ML methodologies, DL techniques frequently employ amalgamated shape descriptors in the form of an additional CNN [62] or an image processing algorithm such as HOG, SURF, or CAT [54,61]. In light of the respective advantages of ML and DL, it is advisable for researchers in the domain of TB-AI to reconsider their endeavours pertaining to medical diagnosis.…”

“…At this stage, when researchers are preparing collections of FOVs for automated analysis, procedures vary. The most common options are: (i) the manual inspection and creation of an image set by an expert [58,63], (ii) auto-focus algorithms [3,29,38], or (iii) successive cropping of the whole slide followed by a filtering stage to remove FOVs void of bacteria [62]. Individual fields of view, or subsections of them, are often additionally cropped into even smaller images, wherein bacteria are present.…”

“…Precision is also known as the predictive positive rate (PPR), which is often accompanied by receiver operating characteristics (ROC) and the area under the ROC curve (AUC). When the context of the task focuses more on accurately identifying positive samples and less on accurately identifying negative samples, it is important to provide an informative analysis using the PPR metric [49,52,62]. These measures assist in the analysis of the trade-off between the true-positive rate (TPR), commonly referred to as sensitivity, and the false-positive rate (FPR), also known as the complement of specificity, across the different decision thresholds of a binary classification model.…”

“…All of the methods listed in this section so far have a key restriction in that they are not fully automated, as they focus on manually cropped, two-digit-sized patches. Zachariou et al, on the other hand, proposed a method to automatically and sequentially slide over the slide to crop FOVs, excluding negative FOVs while keeping all positive ones [62]. They proposed two distinct encoders, one of which was trained to differentiate between bacteria by inferring pixel intensity, while the other was trained to do so by deducing bacterial shape.…”