Eigenloss: Combined PCA-Based Loss Function for Polyp Segmentation

Comprehensive review of publicly available colonoscopic imaging databases for artificial intelligence research: availability, accessibility, and usability

BACKGROUND AND AIMS

Publicly available databases containing colonoscopic imaging data are valuable resources for artificial intelligence (AI) research. Currently, little is known regarding the available number and content of these databases. This review aimed to describe the availability, accessibility, and usability of publicly available colonoscopic imaging databases, focusing on polyp detection, polyp characterization, and quality of colonoscopy.

METHODS

A systematic literature search was performed in MEDLINE and Embase to identify AI studies describing publicly available colonoscopic imaging databases published after 2010. Second, a targeted search using Google's Dataset Search, Google Search, GitHub, and Figshare was done to identify databases directly. Databases were included if they contained data about polyp detection, polyp characterization, or quality of colonoscopy. To assess accessibility of databases, the following categories were defined: open access, open access with barriers, and regulated access. To assess the potential usability of the included databases, essential details of each database were extracted using a checklist derived from the Checklist for Artificial Intelligence in Medical Imaging.

RESULTS

We identified 22 databases with open access, 3 databases with open access with barriers, and 15 databases with regulated access. The 22 open access databases contained 19,463 images and 952 videos. Nineteen of these databases focused on polyp detection, localization, and/or segmentation; 6 on polyp characterization, and 3 on quality of colonoscopy. Only half of these databases have been used by other researcher to develop, train, or benchmark their AI system. Although technical details were in general well reported, important details such as polyp and patient demographics and the annotation process were under-reported in almost all databases.

CONCLUSIONS

This review provides greater insight on public availability of colonoscopic imaging databases for AI research. Incomplete reporting of important details limits the ability of researchers to assess the usability of current databases.

Novel Pixelwise Co-Registered Hematoxylin-Eosin and Multiphoton Microscopy Image Dataset for Human Colon Lesion Diagnosis

Colorectal cancer presents one of the most elevated incidences of cancer worldwide. Colonoscopy relies on histopathology analysis of hematoxylin-eosin (H&E) images of the removed tissue. Novel techniques such as multi-photon microscopy (MPM) show promising results for performing real-time optical biopsies. However, clinicians are not used to this imaging modality and correlation between MPM and H&E information is not clear. The objective of this paper is to describe and make publicly available an extensive dataset of fully co-registered H&E and MPM images that allows the research community to analyze the relationship between MPM and H&E histopathological images and the effect of the semantic gap that prevents clinicians from correctly diagnosing MPM images. The dataset provides a fully scanned tissue images at 10x optical resolution (0.5 µm/px) from 50 samples of lesions obtained by colonoscopies and colectomies. Diagnostics capabilities of TPF and H&E images were compared. Additionally, TPF tiles were virtually stained into H&E images by means of a deep-learning model. A panel of 5 expert pathologists evaluated the different modalities into three classes (healthy, adenoma/hyperplastic, and adenocarcinoma). Results showed that the performance of the pathologists over MPM images was 65% of the H&E performance while the virtual staining method achieved 90%. MPM imaging can provide appropriate information for diagnosing colorectal cancer without the need for H&E staining. However, the existing semantic gap among modalities needs to be corrected.

Unified Focal loss: Generalising Dice and cross entropy-based losses to handle class imbalanced medical image segmentation

Automatic segmentation methods are an important advancement in medical image analysis. Machine learning techniques, and deep neural networks in particular, are the state-of-the-art for most medical image segmentation tasks. Issues with class imbalance pose a significant challenge in medical datasets, with lesions often occupying a considerably smaller volume relative to the background. Loss functions used in the training of deep learning algorithms differ in their robustness to class imbalance, with direct consequences for model convergence. The most commonly used loss functions for segmentation are based on either the cross entropy loss, Dice loss or a combination of the two. We propose the Unified Focal loss, a new hierarchical framework that generalises Dice and cross entropy-based losses for handling class imbalance. We evaluate our proposed loss function on five publicly available, class imbalanced medical imaging datasets: CVC-ClinicDB, Digital Retinal Images for Vessel Extraction (DRIVE), Breast Ultrasound 2017 (BUS2017), Brain Tumour Segmentation 2020 (BraTS20) and Kidney Tumour Segmentation 2019 (KiTS19). We compare our loss function performance against six Dice or cross entropy-based loss functions, across 2D binary, 3D binary and 3D multiclass segmentation tasks, demonstrating that our proposed loss function is robust to class imbalance and consistently outperforms the other loss functions. Source code is available at: https://github.com/mlyg/unified-focal-loss.

•

Loss function choice is crucial for class-imbalanced medical imaging datasets.

•

Understanding the relationship between loss functions is key to inform choice.

•

Unified Focal loss generalises Dice and cross-entropy based loss functions.

•

Unified Focal loss outperforms various Dice and cross-entropy based loss functions.

Automatic Polyp Segmentation in Colonoscopy Images Using a Modified Deep Convolutional Encoder-Decoder Architecture

Colorectal cancer has become the third most commonly diagnosed form of cancer, and has the second highest fatality rate of cancers worldwide. Currently, optical colonoscopy is the preferred tool of choice for the diagnosis of polyps and to avert colorectal cancer. Colon screening is time-consuming and highly operator dependent. In view of this, a computer-aided diagnosis (CAD) method needs to be developed for the automatic segmentation of polyps in colonoscopy images. This paper proposes a modified SegNet Visual Geometry Group-19 (VGG-19), a form of convolutional neural network, as a CAD method for polyp segmentation. The modifications include skip connections, 5 × 5 convolutional filters, and the concatenation of four dilated convolutions applied in parallel form. The CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB databases were used to evaluate the model, and it was found that our proposed polyp segmentation model achieved an accuracy, sensitivity, specificity, precision, mean intersection over union, and dice coefficient of 96.06%, 94.55%, 97.56%, 97.48%, 92.3%, and 95.99%, respectively. These results indicate that our model performs as well as or better than previous schemes in the literature. We believe that this study will offer benefits in terms of the future development of CAD tools for polyp segmentation for colorectal cancer diagnosis and management. In the future, we intend to embed our proposed network into a medical capsule robot for practical usage and try it in a hospital setting with clinicians.

PICCOLO White-Light and Narrow-Band Imaging Colonoscopic Dataset: A Performance Comparative of Models and Datasets

Colorectal cancer is one of the world leading death causes. Fortunately, an early diagnosis allows for effective treatment, increasing the survival rate. Deep learning techniques have shown their utility for increasing the adenoma detection rate at colonoscopy, but a dataset is usually required so the model can automatically learn features that characterize the polyps. In this work, we present the PICCOLO dataset, that comprises 3433 manually annotated images (2131 white-light images 1302 narrow-band images), originated from 76 lesions from 40 patients, which are distributed into training (2203), validation (897) and test (333) sets assuring patient independence between sets. Furthermore, clinical metadata are also provided for each lesion. Four different models, obtained by combining two backbones and two encoder–decoder architectures, are trained with the PICCOLO dataset and other two publicly available datasets for comparison. Results are provided for the test set of each dataset. Models trained with the PICCOLO dataset have a better generalization capacity, as they perform more uniformly along test sets of all datasets, rather than obtaining the best results for its own test set. This dataset is available at the website of the Basque Biobank, so it is expected that it will contribute to the further development of deep learning methods for polyp detection, localisation and classification, which would eventually result in a better and earlier diagnosis of colorectal cancer, hence improving patient outcomes.