Abstract:Background A large training dataset with high-quality annotations is necessary for building an accurate and generalisable deep learning system, which can be difficult and expensive to prepare in medical applications. We present a novel deep-learning-based system, requiring no annotator but weak annotation from a diagnosis report, for accurate and generalisable performance in detecting multiple head disorders from CT scans, including ischaemia, haemorrhage, tumours, and skull fractures. Methods Our system was developed on 104 597 head CT scans from the Chinese PLA General Hospital, with associated textual diagnosis reports. Without expert annotation, we used keyword matching on the reports to automatically generate disorder labels for each scan. The labels were inaccurate because of the unreliable annotator-free strategy and inexact because of scan-level annotation. We proposed RoLo, a novel weakly supervised learning algorithm, with a noise-tolerant mechanism and a multi-instance learning strategy to address these issues. RoLo was tested on retrospective (2357 scans from the Chinese PLA General Hospital), prospective (650 scans from the Chinese PLA General Hospital), cross-centre (1525 scans from the Brain Hospital of Hunan Province), cross-equipment (1484 scans from the Chinese PLA General Hospital), and cross-nation (CQ500 public dataset from India) test datasets. Four radiologists were tested on the prospective test dataset before and after viewing system recommendations to assess whether the system could improve diagnostic performance. Findings The area under the receiver operating characteristic curve for detecting the four disorder types was 0.976 (95% CI 0.976-0.976) for retrospective, 0.975 (0.974-0.976) for prospective, 0.965 (0.964-0.966) for cross-centre, and 0.971 (0.971-0.972) for cross-equipment test datasets, and 0.964 (0.964-0.966) for CQ500 (with only haemorrhage and fracture). The system achieved similar performance to four radiologists and helped to improve sensitivity and specificity by 0.109 (95% CI 0.086-0.131) and 0.022 (0.017-0.026), respectively. Interpretation Without expert annotated data, our system achieved accurate and generalisable performance for head disorder detection. The system improved the diagnostic performance of radiologists. Because of its accuracy and generalisability, our computer-aided diganostic system could be used in clinical practice to improve the accuracy and efficiency of radiologists in different hospitals. (C) 2022 The Author(s). Published by Elsevier Ltd. This is an Open Access article under the CC BY-NC-ND 4.0 license.

Retracted: Diagnosis Labeling With Disease-Specific Characteristics Mining (Retracted Article. See Vol. 103, 2020)

RETRACTED: Diagnosis Labeling with Disease-Specific Characteristics Mining.

Rectified-Linear-Unit-Based Deep Learning for Biomedical Multi-label Data

A Disease Labeler for Chinese Chest X-Ray Report Generation

RETRACTED: Identifying COVID-19-Specific Transcriptomic Biomarkers with Machine Learning Methods (Retracted Article)

Classification Model on Big Data in Medical Diagnosis Based on Semi-Supervised Learning

Multigranularity Label Prediction Model for Automatic International Classification of Diseases Coding in Clinical Text.

Reliability-based cleaning of noisy training labels with inductive conformal prediction in multi-modal biomedical data mining

Multi-label Learning for the Diagnosis of Cancer and Identification of Novel Biomarkers with High-throughput Omics

Language Model Classifier Aligns Better with Physician Word Sensitivity than XGBoost on Readmission Prediction

Feature Selection Method Based on Class Discriminative Degree for Intelligent Medical Diagnosis

Improving Rare Disease Classification Using Imperfect Knowledge Graph

Deep Learning with Weak Annotation from Diagnosis Reports for Detection of Multiple Head Disorders: a Prospective, Multicentre Study

BoMD: Bag of Multi-label Descriptors for Noisy Chest X-ray Classification

An Unsupervised Error Detection Methodology for Detecting Mislabels in Healthcare Analytics

Noisy Label Learning for Large-scale Medical Image Classification

Graph Neural Network Based Multi-Label Hierarchical Classification for Disease Predictions in General Practice.

Hierarchical Knowledge Guided Learning for Real-world Retinal Disease Recognition

Learning with Unsure Data for Medical Image Diagnosis

Clinical multi-label free text classification by exploiting disease label relation

Retracted: Detection of Diseases Using Machine Learning Image Recognition Technology in Artificial Intelligence