METAstasis in LYmph Nodes in oral squamous cell Carcinoma - Histopathology image Dataset AKA METALyNC-HD

METAstasis in LYmph Nodes in oral squamous cell Carcinoma - Histopathology image Dataset, also known as 'METALyNC-HD', was prepared for supporting the development and evaluation of Linux-based AI-driven methodologies for detecting metastatic involvement in lymph nodes. This dataset includes the following subsets: 'Training and Validation dataset' and 'Test dataset', which comprise high-resolution microscopic images and associated diagnostic labels (metastatic or metastasis-free). The 'Training and Validation dataset' contains 16,426 'Metastatic' image patches and 14,192 'Normal' image patches with a size of 256 x 256 pixels, which have been used to train the following five pre-trained deep neural network-based ImageNet models using the transfer learning approach: (i) InceptionV3, (ii) MobileNetV2, (iii) Xception, (iv) ResNet50, and (v) InceptionResNetV2. The Xception model showed the best fit and classification performance. Using the fine-tuned Xception model as a backbone, Automated Lymph Node Metastasis Detection Software (ALNMDS) was developed. ALNMDS showed 82.96% accuracy, 99.8% sensitivity, 80.9% positive predictive value, and 98.77% negative predictive value over the testing dataset that contains 2533 histological images (1024 x 768 pixels) of lymph node metastasis and 999 metastasis-free images (1024 x 768 pixels) of lymph nodes. No patient data has been shared in this dataset. The record is publicly accessible upon request and consideration of the competent authority at the Department of Oral Pathology, Maulana Azad Institute of Dental Sciences, New Delhi, India. The files can only be accessed by following the approval of the competent authority. Legends for the associated supplemental figures: Supplementary File 1 Comparative representation of accuracy graphs, loss graphs, confusion matrix, and AUC-ROC (Area under the Receiver Operating Characteristic Curve) for five fine-tuned deep CNN models during training and validation: InceptionV3, ResNet50, MobileNetV2, Xception, and InceptionResNetV2. Supplemental Figure 2A. Examples of true positive image patches (40x, Hematoxylin and Eosin): (I to VI) Tumor cells in such image patches showed: Cellular and nuclear pleomorphism, nuclear hyperchromasia, increased nucleo-cytoplasmic ratio, and mitotic figures; (VII) Flakes of keratin and tumor cells; (VIII) Small positive tumor patches with a very small deposit in a background of other non-neoplastic cells; (IX) multiple tumor cells showing vesicular nuclei.B. Examples of true negative image patches (40x, Hematoxylin and Eosin): (I to IV) Normal architecture of lymph node showing lymphoid tissue with a thin, reticular network of fibers and few fine endothelial lined blood vessels; mitotic figure; (V) Normal lymph node tissue with plump endothelial cells and pronounced stroma; (VI, VIII, IX) Endothelial cells showing intensely stained nuclei.C. Examples of false-positive image patches (40x, Hematoxylin and Eosin): (I, III, VII) Macrophages with hyperchromatic nuclei pushed to one side of the cell and irregular cellular outline; (II) Plasma cells with hyperchromatic nuclei showing cartwheel appearance, also contained in fibrous connective tissue stroma; (IV) Close arrangement of two cells with hyperchromatic nuclei resembling mitotic figure; (V) Normal lymph node tissue with plump endothelial cells and pronounced stroma, (VI, VIII, IX) Endothelial cells showing intensely stained nuclei.D. Examples of false negative image patches (40x, Hematoxylin and Eosin): (I to II) Cellular and nuclear pleomorphism of tumor cells showing nuclear hyperchromasia, the increased nucleo-cytoplasmic ratio in false negative image patches; (IV to IX) Areas containing a complex pattern of arrangement of neoplastic and non-neoplastic cells with intense staining. Supplemental Figure 3Scatter chart for prediction scores observed with Xception-based model. A. true positive image patches; B. true negative image patches; C. false positive image patches; and D. false negative image patches. Pie charts in each segment show the range of prediction scores for each component of the confusion matrix. Supplemental Figure 4A few examples of the heatmap overlay output of ALNMDS (Automated Lymph Node Metastasis Detection System) showing predictions done by fine-tuned Xception model: (A, B, C) True positive cases showing accurate predictions made by the model; (D, E, F) True negative cases showing complete negative predictions; (G, H) False negative image patches showing areas with normal lymphocytes, endothelial cells and areas showing crushing artifacts being misidentified as positive regions respectively; (I) Hard positive regions being mispredicted as negative regions for metastasis.