Research Presentation Session

RPS 1402a - Artificial intelligence, radiomics and more: part 2

Lectures

1
RPS 1402a - Classification of benign and malignant breast lesions using ultrasound shear wave elastography features: a non-black-box machine learning approach

RPS 1402a - Classification of benign and malignant breast lesions using ultrasound shear wave elastography features: a non-black-box machine learning approach

07:26A. Angelakis, Athens / GR

Purpose:

To tune a non-black-box robust machine learning model on a relatively big, heterogeneous and multicentric data-set including shear wave elastography (SWE) features in order to classify benign and malignant breast lesions achieving high specificity and sensitivity.

Methods and materials:

The data-set consisted of 1,989 breast lesions (995-994, benign-malignant determined by cytopathology or follow-up) coming from 16 European/American and 22 Asian centres with BI-RADS 2-5 scores. Features of the initial data-set were age, palpability, mobility, SWE lesion shape, and homogeneity of the mass. From measures on 3 SWE images we also included the SWE lesion dimension, the maximal and mean SWE values in Q-Box areas on lesion and fatty tissue, and the ratio of mean SWE values on lesion versus fatty tissue. We used feature engineering and tuned a CatBoost classifier of 97 iterators of depth 11.

Results:

The performance of a 10-fold Cross-Validation was: sensitivity: 0.8732, specificity: 0.8983, and ROC-AUC: 0.8858 with 95% CI [0.849 - 0.930]. On an unseen validation dataset, the model’s performance was: sensitivity: 0.8918, specificity: 0.8963, and ROC-AUC: 0.8941. The BI-RADS performance was: sensitivity: 0.9758, specificity: 0.5175, and ROC-AUC: 0.7466, and on the validation data-set was: sensitivity: 0.9837, specificity: 0.5233, and ROC-AUC: 0.7535. The SWE Emax performance was: sensitivity: 0.8490, specificity: 0.8182, and ROC-AUC: 0.8336, and on the validation data-set: sensitivity: 0.8864, specificity: 0.8031, and ROC-AUC: 0.8447.

Conclusion:

In this study, we used a heterogeneous data-set of breast lesions and we tuned a gradient boosted tree classifier trained on SWE features. It achieved high classification scores outperforming conventional approaches (BI-RADS and SWE cut-off value). The model’s result may be considered from radiologists during an examination’s medical report.

Limitations:

n/a

Ethics committee approval

n/a

Funding:

No funding was received for this work.

2
RPS 1402a - Quantitative analysis of contrast-enhanced ultrasound imaging omics in evaluating the efficacy of adriamycin combined with cetuximab in the treatment of triple-negative breast cancer in nude mice

RPS 1402a - Quantitative analysis of contrast-enhanced ultrasound imaging omics in evaluating the efficacy of adriamycin combined with cetuximab in the treatment of triple-negative breast cancer in nude mice

05:47Lei Tang, Shanghai / CN

Purpose:

To evaluate the effect of two-dimensional contrast-enhanced ultrasound (CEUS) on the chemotherapy of triple-negative breast cancer in nude mice and to find out the changes of haemodynamics in the tumours during the treatment.

Methods and materials:

A female BALB/c nude mouse model of triple-negative breast cancer was established by human breast cancer MDA-MB-231 cells. Esaote Mylab90 ultrasound instrument and SonoVue contrast agent were used. Contrast-enhanced ultrasound (CEUS) was performed before administration of adriamycin combined with cetuximab on the 1st, 3rd, and 5th day, and before execution on the 7th day. The images were quantitatively analysed by contrast-enhanced imaging omics. On the basis of delineating the tumour boundary, different ROI regions were extracted respectively (more details are shown in the attached picture).

Results:

16 nude mice in the experimental group completed the 7-day experiment, while 5 in the control group failed to reach the 7-day survival period. All parameters basically reflected the characteristics of increasing with the number of days of treatment, which indicated the normalisation of blood vessels in lesions (more details are shown in the attached picture).

Conclusion:

MTT, TTP, and BI values in the quantitative analysis of two-dimensional contrast-enhanced ultrasound imaging omics showed an upward trend during the treatment of adriamycin combined with cetuximab, which is expected to be helpful in predicting the efficacy of combined therapy for triple-negative breast cancer and will be more suggestive in the prediction of curative effect.

Limitations:

The number of mice used in our experiment is not large enough and there may be selection bias from ultrasound instrument and quantitative analysis software.

Ethics committee approval

Our animal experimental research has been approved by the Animal Ethics Committee of Shanghai Jiaotong University School of Medicine.

Funding:

National Natural Science Foundation of China.

3
RPS 1402a - Radiomic standardisation of breast MRI to predict pathological complete response to neoadjuvant chemotherapy

RPS 1402a - Radiomic standardisation of breast MRI to predict pathological complete response to neoadjuvant chemotherapy

05:51Pia Akl, Lyon / FR

Purpose:

To optimise and standardise breast MRI texture measurements performed as part of a neoadjuvant chemotherapy (NAC) protocol by defining a data preprocessing method before radiomic index extraction, then testing the ability of textural analysis to predict a pathologic complete response (pCR).

Methods and materials:

A clinical dataset of 76 patients, acquired using two scanners and three coils with locally advanced breast tumour diagnosis treated at Institut Curie, were analysed retrospectively. A bias field correction was systematically applied and evaluated through a mean relative difference parameter between left and right ROIs in normal tissues: breast parenchyma (BP) and pectoral muscles (PM). Images were then standardised using the sternum as reference tissue. Tumours were segmented and 48 radiomic features (RF) were extracted from 3D corrected and standardised images using the LIFEx software.

Results:

Relative differences in normal breast and pectoral muscles were significantly reduced (p-value < 0.001 for both BP and PM) using bias field correction. Pathological analysis indicated that 35 (46%) patients were complete responders. The most robust RF for pCR prediction was “GLCM correlation”, showing an AUC under the ROC curve equal to 0.68. In addition, multivariate analysis taking into account different coils and scanners type and molecular subtypes showed no significant difference for this RF.

Conclusion:

This study showed that retrospective bias field correction in breast MRI could improve regional quantitative analyses. This correction was combined with image standardisation on T2 weighted MR images and showed that "GLCM_correlation" is a promising and robust predictor of pCR in breast cancer.

Limitations:

This study should be confirmed on a larger number of patients.

Ethics committee approval

The Institutional Review Committees approved this retrospective study and waived informed consent.

Funding:

No funding was received for this work.

4
RPS 1402a - Radiomics-based MR features in HER2 overexpressing breast cancer receiving neoadjuvant chemotherapy: correlation with pathologic response

RPS 1402a - Radiomics-based MR features in HER2 overexpressing breast cancer receiving neoadjuvant chemotherapy: correlation with pathologic response

05:56A. Bitencourt, Sao Paulo / BR

Purpose:

To use magnetic resonance (MR)-based radiomic features to assess tumour heterogeneity in HER2 overexpressing breast cancer patients receiving neoadjuvant chemotherapy (NAC) and correlate these findings with a pathologic response.

Methods and materials:

This retrospective single-centre study included 311 patients with HER2 overexpressing invasive breast carcinoma who received NAC following pretreatment MRI. Pathologic complete response (pCR) was defined as no residual invasive carcinoma in the breast (ypT0/is) after surgical resection. A breast radiologist performed 3D segmentations of the tumour in the first minute post-contrast sequence using ITK-SNAP software. Enhancement maps were calculated as the percentage increase in signal from the pre-contrast image to the first post-contrast image. Radiomics and statistical analysis were performed using publicly available CERR software and MATLAB.

Results:

Mean tumour size by MRI was 4.7 cm (range: 0.9-14.8 cm). The index tumour presented as mass in 47.3%, non-mass enhancement (NME) in 10.9%, and both mass and NME in 41.8%. Most tumours were multifocal on MRI (65.9%). Overall pCR rate was 62.7% (195/311). 12 radiomics parameters demonstrated significant differences between pCR and non-pCR groups. After ROC and correlation analysis 3 radiomic parameters were retained and advanced to modelling alongside clinical parameters, including lesion type (mass/NME/both), multifocality, size, and nodal status. A robust model was developed utilising coarse decision trees and 5-fold cross-validation. The final model utilised 5 parameters (2 clinical and 3 radiomic) for a diagnostic accuracy of 86.2% (268/311).

Conclusion:

A model including both clinical and radiomics-based MR features can be used to assess tumour heterogeneity and predict pCR after NAC in HER2 overexpressing breast cancer patients.

Limitations:

Retrospective design.

Ethics committee approval

IRB approved.

Funding:

NIH/NCI Cancer Center Support Grant (P30-CA008748) and Breast Cancer Research Foundation.

5
RPS 1402a - Parenchymal radiomics in cone-beam breast CT: comparison with mammography and implication for cancer risk estimation

RPS 1402a - Parenchymal radiomics in cone-beam breast CT: comparison with mammography and implication for cancer risk estimation

05:55Y. Zhu, Tianjin / CN

Purpose:

To investigate the potential advantage of parenchymal radiomics in cone-beam breast CT (CBBCT) compared to mammography (MG) for breast cancer risk estimation.

Methods and materials:

Bilateral CBBCT and MG from 233 women (147 malignant and 86 benign patients who had unilateral disease) were retrospectively collected from a CBBCT clinical trial (NCT01792999) between May 2012 to November 2014. Parenchymal radiomic features were computed from retroareolar region in contralateral normal breast parenchyma. Principal component analysis (PCA) was applied to obtain orthogonal radiomic components. Breast percent density (PD) was evaluated with software. Correlation analysis and linear regression were performed to determine the association between radiomic features and breast PD with increasing levels of risk. Age was also considered as an additional predictor in multivariate models.

Results:

Overall, CBBCT radiomic features demonstrated stronger correlations with breast PD than MG. When dividing population into groups of increasing levels of risk according to breast PD, CBBCT radiomic features appeared to be more discriminative, having regression lines with overall steeper slopes, higher R2 estimates, and lower P values. Linear regression of PCA radiomic features as predictors of PD also demonstrated significantly stronger associations with CBBCT (R2 = 0.375) than with MG (R2 = 0.069). The association was strongest when age was combined (R2 = 0.439).

Conclusion:

Parenchymal radiomic features were more strongly correlated to breast PD in CBBCT than in MG. CBBCT parenchymal radiomics could provide a more accurate characterisation of breast parenchymal patterns, which could ultimately improve breast cancer risk estimation.

Limitations:

Our study cohort was a diagnostic population.

Ethics committee approval

The institutional review board approved this study and written informed consent was obtained from all patients.

Funding:

National Key R&D Program of China (No. 2017YFC0112600), National Natural Science Foundation of China (No. 81571671).

6
RPS 1402a - Artificial intelligence breast cancer risk estimation from CT thorax scans

RPS 1402a - Artificial intelligence breast cancer risk estimation from CT thorax scans

05:32S. De Buck, Leuven / BE

Purpose:

Breast glandularity is associated with breast cancer risk. Systematic scoring of glandularity on CT thorax examinations performed for another clinical reason could find patients at risk, but it is currently cumbersome and not standard practice. Automated scoring would give access to this information. We propose a novel method to automatically segment the breast volume on CT thorax examinations and estimate glandularity.

Methods and materials:

We used an artificial neural network with a U-Net-like architecture to automatically segment the breast region. A retrospective study was conducted on a dataset of 23 postmenopausal women that had a CT thorax examination for another clinical reason. The images were manually segmented to serve as a gold standard. The image/segmentation pairs were randomly separated in a training (15 patients) and a testing (8 patients) set. We computed 2 risk scores in the segmented breast: the glandular fraction (HU>40) and volumetric breast density (VBD), based on the average HU in comparison to pure fat and gland. Segmentation accuracy was validated by Dice overlap between predicted and manually segmented regions. Risk score correlation was assessed by the Pearson coefficient.

Results:

We obtained a Dice overlap score of 0.84±0.03. The Pearson coefficient between predicted and ground-truth values of the glandular fraction and VBD amounted to 0.99 and 0.98 respectively.

Conclusion:

We present a first, novel artificial intelligence-based approach to automatically compute the volume of the breast by a convolutional neural network and determine cancer risk scores based on CT thorax examinations performed for another reason. Results show excellent correspondence between the automated method and expert observers.

Limitations:

Despite a limited population, good results are shown.

Ethics committee approval

A retrospective study on anonymised images, waived by the ethical committee (ZOL).

Funding:

No funding was received for this work.

PEP Subscription Required

This course is only accessible for ESR Premium Education Package subscribers.