Multi-antigen imaging data of skin tissue samples from melanoma and benign nevi

<h2><strong>Data</strong></h2> <p>Multi-antigen imaging data with 182 skin tissue samples:</p> <table> <tbody> <tr> <td>Condition</td> <td>PFS</td> <td>Number of patients</td> <td>Number of MELC images</td> </tr> <tr> <td>Melanoma</td> <td>PFS &gt;=5</td> <td>14</td> <td>97</td> </tr> <tr> <td>Melanoma</td> <td>PFS &lt; 5</td> <td>8</td> <td>15</td> </tr> <tr> <td>Melanoma</td> <td>NA</td> <td>5</td> <td>13</td> </tr> <tr> <td>Benign nevi</td> <td>-</td> <td>12</td> <td>57</td> </tr> </tbody> </table> <p>Patients were treated at the University Hospital Erlangen. Each sample contains around 50 protein channels, each of resolution 512 X 512 pixels. The data were generated using multi-epitope ligand cartography (MELC) (https://doi.org/10.1038/nbt1250, https://doi.org/10.1007/3-540-36459-5_8).</p> <h2>Metadata</h2> <p>For each MELC image, the table contains the corresponding section and patient ID, sex, age, and which group (melanoma or nevus) it belongs to. For most of the melanoma samples, the data contains the information, whether the patient had progression free survival (PFS) of 5 years, the tumor thickness, whether the tumor is ulcerated and its coarse location (0: thoraric, 1: abdominal, 2: back, 3: arm, 4: leg) and side (right: -1, left: 1). Furthermore, we annotate if we could produce a high-quality segmentation result.</p> <h2>Model weights</h2> <p>We provide the pre-trained model as well as the fine-tuned models. Please be aware that we initalize the model before pre-training with the weights provided here https://github.com/ozanciga/self-supervised-histopathology/releases/tag/tenpercent (Ciga, O., Xu, T., and Martel, A. L. (2022). Self supervised contrastive learning for digital histopathology. Machine Learning with Applications 7, 100198).</p> <h2>Anndata file</h2> <p>We provide an anndata file with cell-level protein abundance and assigned cell types.</p>