Automatic Detection of Virus Infection Patterns in Foci Images Using Switchable Convolutions

Detection of small objects with undefined shapes, such as color-stained one or a group of virus-infected cells, i.e. foci, in microscopy images, is challenging due to their size and scale variation. Quantifying the foci numbers manually using light microscope images is a laborious task. On the other hand, state-of-the-art machine learning methods often fail to detect uneven foci due to overfitting because of the unavailability of the required number of training images. This study proposes a Scale-Invariant Object Detection method based on Switchable Atrous Convolution (SIOD_SAC) for detecting foci. The proposed method applies a Switchable Atrous Convolution (SAC) module to extract multi-scale features and maintain scale invariance in object detection. The proposed SAC module adaptively selects different dilation rates to capture features at multiple scales, facilitating the detection of objects of different sizes. Due to the unavailability of a foci image dataset for virus patch detection, we introduce a dataset containing 149 foci images with virus patches of different shapes and sizes. Here, we utilized a dataset comprising 149 images of various shaped and sized foci. The microscopic images in the dataset are obtained from the hepatitis C virus focus-forming assay. We evaluate the proposed method on the introduced dataset for small object detection, and our results show that the proposed method achieves state-of-the-art performance in terms of mean Average Precision (mAP). Specifically, the proposed SIOD_SAC method achieves an mAP of 24.57% for cell patch detection and 76.19% for FFU (focus forming units) detection at the threshold of 25 IOU, outperforming the current state-of-the-art by a significant margin. The dataset can be obtained through https://forms.gle/fbhJrBt9akJkkNz49 . The code is available at https://github.com/asAmrita/Virus_Patch_Detection .