Predicting Unnecessary Nodule Biopsies from a Small, Unbalanced, and Pathologically Proven Dataset by Transfer Learning

Leveraging prior knowledge in machine intelligence to improve lesion diagnosis for early cancer detection

BACKGROUND

Experts' interpretations of medical images for lesion diagnosis may not always align with the underlying in vivo tissue pathology and, therefore, cannot be considered the definitive truth regarding malignancy or benignity. While current machine learning (ML) models in medical imaging can replicate expert interpretations, their results may also diverge from the actual ground truth.

PURPOSE

This study investigates various factors contributing to these discrepancies and proposes solutions.

METHODS

The central idea of the proposed solution is to integrate prior knowledge into ML models to enhance the characterization of in vivo tissues. The incorporation of prior knowledge into decision-making is task-specific, tailored to the data acquired for that task. This central idea was tested on the diagnosis of lesions using low dose computed tomography (LdCT) for early cancer detection, particularly focusing on more challenging, ambiguous or indeterminate lesions (IDLs) as classified by experts. One key piece of prior knowledge involves CT x-ray energy spectrum, where different energies interact with in vivo tissues within a lesion, producing variable but reproducible image contrasts that encapsulate biological information. Typically, CT imaging devices use only the high-energy portion of this spectrum for data acquisition; however, this study considers the full spectrum for lesion diagnostics. Another critical aspect of prior knowledge includes the functional or dynamic properties of in vivo tissues, such as elasticity, which can indicate pathological conditions. Instead of relying solely on abstract image features as current ML models do, this study extracts these tissue pathological characteristics from the image contrast variations.

RESULTS

The method was tested on LdCT images of four sets of IDLs, including pulmonary nodules and colorectal polyps, with pathological reports serving as the ground truth for malignancy or benignity. The method achieved an area under the receiver operating characteristic curve (AUC) of 0.98 ± 0.03, demonstrating a significant improvement over existing state-of-the-art ML models, which typically have AUCs in the 0.70 range.

CONCLUSION

Leveraging prior knowledge in machine intelligence can enhance lesion diagnosis, resolve the ambiguity of IDLs interpreted by experts, and improve the effectiveness of LdCT screening for early-stage cancers.

Automated prediction of early spontaneous miscarriage based on the analyzing ultrasonographic gestational sac imaging by the convolutional neural network: a case-control and cohort study

Background

It is challenging to predict the outcome of the pregnancy when fetal heart activity is detected in early pregnancy. However, an accurate prediction is of importance for obstetricians as it helps to provide appropriate consultancy and determine the frequency of ultrasound examinations. The purpose of this study was to investigate the role of the convolutional neural network (CNN) in the prediction of spontaneous miscarriage risk through the analysis of early ultrasound gestational sac images.

Methods

A total of 2196 ultrasound images from 1098 women with early singleton pregnancies of gestational age between 6 and 8 weeks were used for training a CNN for the prediction of the miscarriage in the retrospective study. The patients who had positive fetal cardiac activity on their first ultrasound but then experienced a miscarriage were enrolled. The control group was randomly selected in the same database from the fetuses confirmed to be normal during follow-up. Diagnostic performance of the algorithm was validated and tested in two separate test sets of 136 patients with 272 images, respectively. Performance in prediction of the miscarriage was compared between the CNN and the manual measurement of ultrasound characteristics in the prospective study.

Results

The accuracy of the predictive model was 80.32% and 78.1% in the retrospective and prospective study, respectively. The area under the receiver operating characteristic curve (AUC) for classification was 0.857 (95% confidence interval [CI], 0.793–0.922) in the retrospective study and 0.885 (95%CI, 0.846–0.925) in the prospective study, respectively. Correspondingly, the predictive power of the CNN was higher compared with manual ultrasound characteristics, for which the AUCs of the crown-rump length combined with fetal heart rate was 0.687 (95%CI, 0.587–0.775).

Conclusions

The CNN model showed high accuracy for predicting miscarriage through the analysis of early pregnancy ultrasound images and achieved better performance than that of manual measurement.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12884-022-04936-0.

Performance Evaluation of a Deep Learning System for Differential Diagnosis of Lung Cancer With Conventional CT and FDG PET/CT Using Transfer Learning and Metadata

PURPOSE

We aimed to evaluate the performance of a deep learning system for differential diagnosis of lung cancer with conventional CT and FDG PET/CT using transfer learning (TL) and metadata.

METHODS

A total of 359 patients with a lung mass or nodule who underwent noncontrast chest CT and FDG PET/CT prior to treatment were enrolled retrospectively. All pulmonary lesions were classified by pathology (257 malignant, 102 benign). Deep learning classification models based on ResNet-18 were developed using the pretrained weights obtained from ImageNet data set. We propose a deep TL model for differential diagnosis of lung cancer using CT imaging data and metadata with SUVmax and lesion size derived from PET/CT. The area under the receiver operating characteristic curve (AUC) of the deep learning model was measured as a performance metric and verified by 5-fold cross-validation.

RESULTS

The performance metrics of the conventional CT model were generally better than those of the CT of PET/CT model. Introducing metadata with SUVmax and lesion size derived from PET/CT into baseline CT models improved the diagnostic performance of the CT of PET/CT model (AUC = 0.837 vs 0.762) and the conventional CT model (AUC = 0.877 vs 0.817).

CONCLUSIONS

Deep TL models with CT imaging data provide good diagnostic performance for lung cancer, and the conventional CT model showed overall better performance than the CT of PET/CT model. Metadata information derived from PET/CT can improve the performance of deep learning systems.