Project results short video series: Human model in-vivo wide-field DB & algorithm
As part of the commitment of the PICCOLO project to make openly accessible scientific data, a Human model in-vivo Wide-field database has been generated for its publication. The dataset consists of white light/NBI images from clinical colonoscopy videos acquired in 48 human patients. For each lesion, the number of polyps of interest, their size and classification, preliminary diagnosis, literal diagnosis and histological stratification are given. In all, 3433 images have been manually annotated to delimitate the polyp area.
With the algorithm automatic segmentation applied to polyps in colonoscopic images we improve the detection of them by highlighting their contour. An automatic polyp segmentation algorithm has been implemented using a Fully Convolutional Network (FCN) and transfer learning between two publicly available medical images databases: CVC-EndoSceneStill and CVC-VideoClinicDB.
For the segmentation algorithm, the first approach used a U-Net based architecture and finetune using different datasets. We have also analyzed the influence of data augmentation transformations on the performance of deep learning models for polyp segmentation, as well as the different loss functions, to define the eigenloss, a PCA-based loss function. With the PICCOLO Human model in-vivo Wide-field dataset, different models based on different backbones and encoder-decoder architecture have been trained, showing the utility of the dataset for polyp detection, localization and segmentation.
The openly accessible dataset is at the scientific community´s disposal for the training of clinicians, or for their use in algorithms for the detection and/or classification of colorectal cancer. The terms and details of these outputs are and will be openly accessible through publications in scientific journals at the scientific community´s disposal.