BCData: A Large-Scale Dataset and Benchmark for Cell Detection and Counting

Deep Learning-Inferred Multiplex ImmunoFluorescence for IHC Image Quantification

“…1) We evaluate all models on our internal test set which includes 600 images of size 512 × 512 and 40x magnification from bladder carcinoma and non-small cell lung carcinoma slides. 2) We randomly selected and segmented 41 images of size 640 × 640 from recently released BCDataset (20) which contains Ki67 stained sections of breast carcinoma with Ki67+ and Ki67-cell centroid annotations (targeting cell detection as opposed to cell instance segmentation task). We split these tiles into 164 images of size 512 × 512; the test set varies widely in the density of tumor cells and the Ki67 index.…”

“…(d) As mentioned earlier, DeepLIIF generalizes across different tissue types and imaging platforms. Two example images from the BC Dataset (20) along with the generated modalities and classified segmentation masks are shown in the top rows where the ground-truth mask and segmentation masks of five state-of-the-art models are shown in the second row. The mean IOU and Pixel Accuracy are given for each generated mask.…”

“…Application of DeepLIIF to the BC Dataset (20) resulted in a pixel accuracy of 84.00%, IOU of 36.70%, and Dice Score of 50.59%, and outperformed Mask_RCNN with pixel accuracy of 83.47%, IOU of 33.36%, and Dice Score of 46.64%, UNet ++ with pixel accuracy of 81.96%, IOU of 34.32%, and Dice Score of 47.88%, LinkNet with pixel accuracy of 82.59%, IOU of 35.05%, and Dice Score of 48.56%, and FPN with pixel accuracy of 80.64%, IOU of 34.25%, and Dice Score of 47.47%, while maintaining lower standard deviation on all metrics.…”

“…However, since DeepLIIF is capable of separating the touching nuclei, we also performed a cell-level analysis of DeepLIIF against cell centroid detection approaches. U-CSRNet (20), for example, detects and classifies cells without performing cell instance segmentation. Most of these approaches use crowd counting techniques to find cell centroids.…”

“…The major hurdle in evaluating these techniques is the variance in detected cell centroids. We trained FCRN_A (18), FCRN_B (18), Deeplab_Xeption (30), SC_CNN (19), CSR-Net (31), U-CSRNet (20) using our training set (the centroids of our individual cell segmentation masks are used as detection masks). Most of these approaches failed in detecting and classifying cells on the BCData testing set, and the rest detected centroids far from the ground-truth centroids.…”