Dataset Information

Unleashing the potential of digital pathology data by training computer-aided diagnosis models without human annotations.

ABSTRACT: The digitalization of clinical workflows and the increasing performance of deep learning algorithms are paving the way towards new methods for tackling cancer diagnosis. However, the availability of medical specialists to annotate digitized images and free-text diagnostic reports does not scale with the need for large datasets required to train robust computer-aided diagnosis methods that can target the high variability of clinical cases and data produced. This work proposes and evaluates an approach to eliminate the need for manual annotations to train computer-aided diagnosis tools in digital pathology. The approach includes two components, to automatically extract semantically meaningful concepts from diagnostic reports and use them as weak labels to train convolutional neural networks (CNNs) for histopathology diagnosis. The approach is trained (through 10-fold cross-validation) on 3'769 clinical images and reports, provided by two hospitals and tested on over 11'000 images from private and publicly available datasets. The CNN, trained with automatically generated labels, is compared with the same architecture trained with manual labels. Results show that combining text analysis and end-to-end deep neural networks allows building computer-aided diagnosis tools that reach solid performance (micro-accuracy = 0.908 at image-level) based only on existing clinical data without the need for manual annotations.

SUBMITTER: Marini N

PROVIDER: S-EPMC9307641 | biostudies-literature | 2022 Jul

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Unleashing the potential of digital pathology data by training computer-aided diagnosis models without human annotations.

Marini Niccolò N Marchesin Stefano S Otálora Sebastian S Wodzinski Marek M Caputo Alessandro A van Rijthoven Mart M Aswolinskiy Witali W Bokhorst John-Melle JM Podareanu Damian D Petters Edyta E Boytcheva Svetla S Buttafuoco Genziana G Vatrano Simona S Fraggetta Filippo F van der Laak Jeroen J Agosti Maristella M Ciompi Francesco F Silvello Gianmaria G Muller Henning H Atzori Manfredo M

NPJ digital medicine 20220722 1

The digitalization of clinical workflows and the increasing performance of deep learning algorithms are paving the way towards new methods for tackling cancer diagnosis. However, the availability of medical specialists to annotate digitized images and free-text diagnostic reports does not scale with the need for large datasets required to train robust computer-aided diagnosis methods that can target the high variability of clinical cases and data produced. This work proposes and evaluates an app ...[more]

PMID: 35869179

Similar Datasets

Project description:ImportanceAfter the US Food and Drug Administration (FDA) approved computer-aided detection (CAD) for mammography in 1998, and the Centers for Medicare and Medicaid Services (CMS) provided increased payment in 2002, CAD technology disseminated rapidly. Despite sparse evidence that CAD improves accuracy of mammographic interpretations and costs over $400 million a year, CAD is currently used for most screening mammograms in the United States.ObjectiveTo measure performance of digital screening mammography with and without CAD in US community practice.Design, setting, and participantsWe compared the accuracy of digital screening mammography interpreted with (n = 495 818) vs without (n = 129 807) CAD from 2003 through 2009 in 323 973 women. Mammograms were interpreted by 271 radiologists from 66 facilities in the Breast Cancer Surveillance Consortium. Linkage with tumor registries identified 3159 breast cancers in 323 973 women within 1 year of the screening.Main outcomes and measuresMammography performance (sensitivity, specificity, and screen-detected and interval cancers per 1000 women) was modeled using logistic regression with radiologist-specific random effects to account for correlation among examinations interpreted by the same radiologist, adjusting for patient age, race/ethnicity, time since prior mammogram, examination year, and registry. Conditional logistic regression was used to compare performance among 107 radiologists who interpreted mammograms both with and without CAD.ResultsScreening performance was not improved with CAD on any metric assessed. Mammography sensitivity was 85.3% (95% CI, 83.6%-86.9%) with and 87.3% (95% CI, 84.5%-89.7%) without CAD. Specificity was 91.6% (95% CI, 91.0%-92.2%) with and 91.4% (95% CI, 90.6%-92.0%) without CAD. There was no difference in cancer detection rate (4.1 in 1000 women screened with and without CAD). Computer-aided detection did not improve intraradiologist performance. Sensitivity was significantly decreased for mammograms interpreted with vs without CAD in the subset of radiologists who interpreted both with and without CAD (odds ratio, 0.53; 95% CI, 0.29-0.97).Conclusions and relevanceComputer-aided detection does not improve diagnostic accuracy of mammography. These results suggest that insurers pay more for CAD with no established benefit to women.

Project description:ObjectiveTo test the performance of an artificial intelligence-based computer-aided diagnosis (AI-CAD) designed for full-field digital mammography (FFDM) when applied to synthetic mammography (SM).Materials and methodsWe analyzed 501 women (mean age, 57 ± 11 years) who underwent preoperative mammography and breast cancer surgery. This cohort consisted of 1002 breasts, comprising 517 with cancer and 485 without. All patients underwent digital breast tomosynthesis (DBT) and FFDM during the preoperative workup. The SM is routinely reconstructed using DBT. Commercial AI-CAD (Lunit Insight MMG, version 1.1.7.2) was retrospectively applied to SM and FFDM to calculate the abnormality scores for each breast. The median abnormality scores were compared for the 517 breasts with cancer using the Wilcoxon signed-rank test. Calibration curves of abnormality scores were evaluated. The discrimination performance was analyzed using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity using a 10% preset threshold. Sensitivity and specificity were further analyzed according to the mammographic and pathological characteristics. The results of SM and FFDM were compared.ResultsAI-CAD demonstrated a significantly lower median abnormality score (71% vs. 96%, P < 0.001) and poorer calibration performance for SM than for FFDM. SM exhibited lower sensitivity (76.2% vs. 82.8%, P < 0.001), higher specificity (95.5% vs. 91.8%, P < 0.001), and comparable AUC (0.86 vs. 0.87, P = 0.127) than FFDM. SM showed lower sensitivity than FFDM in asymptomatic breasts, dense breasts, ductal carcinoma in situ, T1, N0, and hormone receptor-positive/human epidermal growth factor receptor 2-negative cancers but showed higher specificity in non-cancerous dense breasts.ConclusionAI-CAD showed lower abnormality scores and reduced calibration performance for SM than for FFDM. Furthermore, the 10% preset threshold resulted in different discrimination performances for the SM. Given these limitations, off-label application of the current AI-CAD to SM should be avoided.

Project description:PurposeEffective diagnosis of tuberculosis (TB) relies on accurate interpretation of radiological patterns found in a chest radiograph (CXR). Lack of skilled radiologists and other resources, especially in developing countries, hinders its efficient diagnosis. Computer-aided diagnosis (CAD) methods provide second opinion to the radiologists for their findings and thereby assist in better diagnosis of cancer and other diseases including TB. However, existing CAD methods for TB are based on the extraction of textural features from manually or semi-automatically segmented CXRs. These methods are prone to errors and cannot be implemented in X-ray machines for automated classification.MethodsGabor, Gist, histogram of oriented gradients (HOG), and pyramid histogram of oriented gradients (PHOG) features extracted from the whole image can be implemented into existing X-ray machines to discriminate between TB and non-TB CXRs in an automated manner. Localized features were extracted for the above methods using various parameters, such as frequency range, blocks and region of interest. The performance of these features was evaluated against textural features. Two digital CXR image datasets (8-bit DA and 14-bit DB) were used for evaluating the performance of these features.ResultsGist (accuracy 94.2% for DA, 86.0% for DB) and PHOG (accuracy 92.3% for DA, 92.0% for DB) features provided better results for both the datasets. These features were implemented to develop a MATLAB toolbox, TB-Xpredict, which is freely available for academic use at http://sourceforge.net/projects/tbxpredict/. This toolbox provides both automated training and prediction modules and does not require expertise in image processing for operation.ConclusionSince the features used in TB-Xpredict do not require segmentation, the toolbox can easily be implemented in X-ray machines. This toolbox can effectively be used for the mass screening of TB in high-burden areas with improved efficiency.

Project description:BackgroundComputer-aided diagnosis (CADx) software that provides a second opinion has been widely used to assist physicians with various tasks. In dermatology, however, CADx has been mostly limited to melanoma or melanocytic skin cancer diagnosis. The frequency of non-melanocytic skin cancers and the accessibility of regular digital macrographs have raised interest in developing CADx for broader applications.ObjectivesTo investigate the feasibility of using CADx to diagnose both melanocytic and non-melanocytic skin lesions based on conventional digital photographic images.MethodsThis study was approved by an institutional review board, and the requirement to obtain informed consent was waived. In total, 769 conventional photographs of melanocytic and non-melanocytic skin lesions were retrospectively reviewed and used to develop a CADx system. Conventional and new color-related image features were developed to classify the lesions as benign or malignant using support vector machines (SVMs). The performance of CADx was compared with that of dermatologists.ResultsThe clinicians' overall sensitivity, specificity, and accuracy were 83.33%, 85.88%, and 85.31%, respectively. New color correlation and principal component analysis (PCA) features improved the classification ability of the baseline CADx (p = 0.001). The estimated area under the receiver operating characteristic (ROC) curve (Az) of the proposed CADx system was 0.949, with a sensitivity and specificity of 85.63% and 87.65%, respectively, and a maximum accuracy of 90.64%.ConclusionsWe have developed an effective CADx system to classify both melanocytic and non-melanocytic skin lesions using conventional digital macrographs. The system's performance was similar to that of dermatologists at our institute. Through improved feature extraction and SVM analysis, we found that conventional digital macrographs were feasible for providing useful information for CADx applications. The new color-related features significantly improved CADx applications for skin cancer.

Dataset Information

Unleashing the potential of digital pathology data by training computer-aided diagnosis models without human annotations.

Publications

Unleashing the potential of digital pathology data by training computer-aided diagnosis models without human annotations.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets