ABSTRACT
Since the COVID-19 pandemic, several research studies have proposed Deep Learning (DL)-based automated COVID-19 detection, reporting high cross-validation accuracy when classifying COVID-19 patients from normal or other common Pneumonia. Although the reported outcomes are very high in most cases, these results were obtained without an independent test set from a separate data source(s). DL models are likely to overfit training data distribution when independent test sets are not utilized or are prone to learn dataset-specific artifacts rather than the actual disease characteristics and underlying pathology. This study aims to assess the promise of such DL methods and datasets by investigating the key challenges and issues by examining the compositions of the available public image datasets and designing different experimental setups. A convolutional neural network-based network, called CVR-Net (COVID-19 Recognition Network), has been proposed for conducting comprehensive experiments to validate our hypothesis. The presented end-to-end CVR-Net is a multi-scale-multi-encoder ensemble model that aggregates the outputs from two different encoders and their different scales to convey the final prediction probability. Three different classification tasks, such as 2-, 3-, 4-classes, are designed where the train-test datasets are from the single, multiple, and independent sources. The obtained binary classification accuracy is 99.8% for a single train-test data source, where the accuracies fall to 98.4% and 88.7% when multiple and independent train-test data sources are utilized. Similar outcomes are noticed in multi-class categorization tasks for single, multiple, and independent data sources, highlighting the challenges in developing DL models with the existing public datasets without an independent test set from a separate dataset. Such a result concludes a requirement for a better-designed dataset for developing DL tools applicable in actual clinical settings. The dataset should have an independent test set; for a single machine or hospital source, have a more balanced set of images for all the prediction classes; and have a balanced dataset from several hospitals and demography. Our source codes and model are publicly available for the research community for further improvements.
ABSTRACT
BACKGROUND AND OBJECTIVE: In modern ophthalmology, automated Computer-aided Screening Tools (CSTs) are crucial non-intrusive diagnosis methods, where an accurate segmentation of Optic Disc (OD) and localization of OD and Fovea centers are substantial integral parts. However, designing such an automated tool remains challenging due to small dataset sizes, inconsistency in spatial, texture, and shape information of the OD and Fovea, and the presence of different artifacts. METHODS: This article proposes an end-to-end encoder-decoder network, named DRNet, for the segmentation and localization of OD and Fovea centers. In our DRNet, we propose a skip connection, named residual skip connection, for compensating the spatial information lost due to pooling in the encoder. Unlike the earlier skip connection in the UNet, the proposed skip connection does not directly concatenate low-level feature maps from the encoder's beginning layers with the corresponding same scale decoder. We validate DRNet using different publicly available datasets, such as IDRiD, RIMONE, DRISHTI-GS, and DRIVE for OD segmentation; IDRiD and HRF for OD center localization; and IDRiD for Fovea center localization. RESULTS: The proposed DRNet, for OD segmentation, achieves mean Intersection over Union (mIoU) of 0.845, 0.901, 0.933, and 0.920 for IDRiD, RIMONE, DRISHTI-GS, and DRIVE, respectively. Our OD segmentation result, in terms of mIoU, outperforms the state-of-the-art results for IDRiD and DRIVE datasets, whereas it outperforms state-of-the-art results concerning mean sensitivity for RIMONE and DRISHTI-GS datasets. The DRNet localizes the OD center with mean Euclidean Distance (mED) of 20.23 and 13.34 pixels, respectively, for IDRiD and HRF datasets; it outperforms the state-of-the-art by 4.62 pixels for IDRiD dataset. The DRNet also successfully localizes the Fovea center with mED of 41.87 pixels for the IDRiD dataset, outperforming the state-of-the-art by 1.59 pixels for the same dataset. CONCLUSION: As the proposed DRNet exhibits excellent performance even with limited training data and without intermediate intervention, it can be employed to design a better-CST system to screen retinal images. Our source codes, trained models, and ground-truth heatmaps for OD and Fovea center localization will be made publicly available upon publication at GitHub.1.
