RPS 905 - Artificial intelligence and machine learning for ultrasound

Purpose:

To evaluate and compare the development of foetal lungs with pregnancy complications and normal pregnancy during different gestational weeks using ultrasound-based radiomics technology.

Methods and materials:

A total of 548 foetal lung ultrasound images of 491 single pregnant women were obtained during routine ultrasound examinations between 28 and 41 weeks of gestation before birth. Ultrasound-based radiomics technology was used to extract foetal lung image features. A standard machine-learning model was composed of feature transformation and a regression model was used to evaluate the relationship between texture features, pregnancy complications, and gestational age.

Results:

Foetal lung ultrasound images were divided into four groups: GDM group, GDM + pre-eclampsia group, pre-eclampsia group, and normal group. The accuracy of foetal lung texture analysis in estimating complications of different pregnancies at different gestational weeks was 80.1%-92.9%. The accuracy of the gestational age prediction model established by foetal lung image texture analysis of the normal and complication groups was 61.5–82.1%.

Conclusion:

Ultrasound-based radiomics technology is a noninvasive way to assess foetal lung development during different pregnancy complications. The foetal lung maturity prediction model will be of great help in the assessment of foetal lung development during different pregnancy complications, for monitoring the medication of disease, and the choice of termination time.

Limitations:

By expanding the sample size, the stability and accuracy of the model will be improved. The GDM groups were not subdivided according to the severity of diabetes mellitus, further emphasising the need to test and verify the differences between examiners and machines.

Ethics committee approval

n/a

Funding:

This work was supported by the National Natural Science Foundation of China (Grant 61771143 and 61871135) and the Science and Technology Commission of Shanghai Municipality (Grant 18511102904 and 17411953400).

Purpose:

To explore the use of parametric imaging as an imaging ensemble method to analyse breast contrast ultrasound images to diagnose or predict luminal A breast cancer in the early stage.

Methods and materials:

189 patients with breast cancer who underwent a contrast-enhanced ultrasound and obtained pathological findings before surgery were enrolled in this retrospective analysis. After extracting the TIC curve of each pixel in the ROI region of images, the relevant parameters PI, AUC, MTT, WiR, WoR, TTP, and RT in the curve were used as independent parameters for two-dimensional imaging. Each parameter was unified between 0 and 255 grey-levels to generate a parametric image grey-scale image. By transforming the index map and the palette axis, the three fluxes were superimposed to generate a parametric imaging colour map.

Results:

There were 46 luminal A breast cancers and a total of 644 parameter imaging features. After the dimension reduction was selected and the features of repetition or the same meaning were eliminated, 3 effective features were finally selected. After the SVM output the probability value, RT_E2 had an AUC of 0.677, a sensitivity of 83.9%, a specificity of 71.7%, and an accuracy of 81.0%. The AUC, sensitivity, specificity, and accuracy of RT_Hm2 were 0.650, 77.4%, 70.3%, and 71.4%, respectively. Those of RT_Et2 reached 0.665, 63.6%, 69.6%, and 65.1%, respectively.

Conclusion:

Parametric imaging features, combined with time and distribution characteristics, provide more comprehensive diagnostic information and offer the possibility of predicting luminal A breast cancer.

Limitations:

The design of this study has limitations. The sample was biased for the different incidence rates.

Ethics committee approval

All patients signed written informed consent.

Funding:

No funding was received for this work.

Purpose:

To predict the presence of placenta adhesion disorder (PAD) in patients with placenta previa using ultrasound (US) derived texture analysis (TA) features and a machine learning (ML) analysis.

Methods and materials:

53 patients (n=36 without PAS and n= 17 with PAS) with placenta previa who underwent endo-vaginal US examination for suspicion of PAD were retrospectively selected. 2D circle ROI was placed over homogeneous placental tissue. ROIs and corresponding images were then imported on a dedicated software (3D slicer, heterogeneity CAD) to extract first, second, and higher order TA features. ML analysis was subsequently run to identify the best performing method to correctly classify instances.

Results:

A total of 688 TA features were extracted. 299 TA features showed an intraclass correlation coefficient values higher or equal to 0.75. No features showed low variance using a threshold of 0.0. Highly (r>0.8) correlated features were excluded, with 14 TA features further selected. Using these features with a bagged J48 algorithm, with each bag containing 70% of cases and 50 iterations, an accuracy of 82% was obtained.

Conclusion:

ML analysis using TA featured extracted from US images may be useful to accurately identify PAS in patients with placenta previa.

Limitations:

Limitations are related to a relatively small sample and a retrospective design.

Ethics committee approval

IRB approved, consent waived.

Funding:

No funding was received for this work.

Purpose:

To propose a transfer learning (TL) radiomics model that efficiently combines the information from grey-scale and elastogram ultrasound images for accurate liver fibrosis grading.

Methods and materials:

Totally, 466 patients undergoing partial hepatectomy were enrolled, including 401 with chronic hepatitis B and 65 without fibrosis pathologically. All patients received elastography and got liver stiffness measurement (LSM) 2-3 days before surgery. We proposed a deep convolutional neural network by TL to analyse images of grey-scale modality (GM) and elastogram modality (EM). The TL process was used for liver fibrosis classification by Inception-V3 network which pretrained on ImageNet. The diagnostic performance of TL and non-TL was compared. The value of single modalities, including GM and EM alone, and multimodalities, including GM+LSM and GM+EM, was evaluated and compared with that of LSM and serological indexes. Receiver operating characteristic curve analysis was performed to calculate the optimal area under the curve (AUC) for classifying fibrosis of S4, ≥S3, and ≥S2.

Results:

TL in GM and EM demonstrated higher diagnostic accuracy than non-TL, with significantly higher AUCs (all P < .01). Single-modal GM and EM both performed better than LSM and serum indexes (all P < .001). Multimodal GM+EM was the most accurate prediction model (AUCs are 0.950, 0.932, and 0.930 for classifying S4, ≥S3, and ≥S2, respectively) compared with GM+LSM, GM and EM alone, LSM, and biomarkers (all P < .05).

Conclusion:

Liver fibrosis can be staged by a transfer learning modal based on the combination of grey-scale and elastogram ultrasound images, with excellent performance.

Limitations:

A single-centre study.

Ethics committee approval

n/a

Funding:

This study received funding from National Natural Science Foundation of China (grant number 81571675, 81873897, and 61471125).

Purpose:

Prenatal imaging is barely performed in resource-limited settings. This is mainly caused by a severe shortage of ultrasound devices and trained sonographers capable of acquiring and interpreting ultrasound images. We investigated if midwives in resource-limited settings are able to perform a predefined ultrasound acquisition protocol. This acquisition protocol consists of six predefined sweeps with the transducer over the abdomen of the pregnant woman. In previous research, we developed deep learning algorithms that can automatically interpret prenatal ultrasound images which were acquired using this protocol.

Methods and materials:

A workshop was given to 5 midwives at St. Luke’s Catholic Hospital in Wolisso, Ethiopia. The midwives had no prior knowledge of ultrasound. The workshop consisted of a presentation to introduce ultrasound imaging and the acquisition protocol. The midwives practised the acquisition protocol on each other to become familiar with the protocol. The midwives then went to the prenatal department to acquire the protocol from pregnant women.

Results:

The midwives successfully acquired the protocol from all 72 pregnant women that visited the hospital. The protocol was acquired using the MicrUs EXT-1H (Telemed ultrasound medical systems, Vilnius, Lithuania). All midwives were able to perform the protocol within two hours of training.

Conclusion:

The predefined ultrasound acquisition protocol can be taught within two hours to midwives without prior knowledge of ultrasound. Combining this protocol with algorithms that automatically interpret the ultrasound data has the potential to make implementation of prenatal screening in resource-limited settings much faster and easier, since it would avoid the need to train sonographers.

Limitations:

This feasibility study only included five midwives from one city.

Ethics committee approval

This study was approved by the local ethics committee. Every woman in this study signed a written informed consent.

Funding:

No funding was received for this work.

Purpose:

To determine the accuracy of a CAD system to classify breast masses on ultrasound according to BI-RADS lexicon 5th edition

Methods and materials:

83 breast masses consecutively referred for biopsy (31 malignant) were included. A 15 years experienced radiologist classified the masses according to criteria proposed by the BI-RADS. For classification, the B-Mode findings associated with the strain elastography findings were considered.
The radiologist final classification was compared to that obtained by a semi-automated CAD system adopting similar classification criteria as proposed by the BI-RADS lexicon.
To perform the CAD system classification, the same observer manually delimited the masses at the B-Mode images. The CAD classification system consists of 3 steps:
1) Use of machine learning to classify masses in B-mode.
2) Quantitative classification of masses by strain elastography.
3) Integration of 1 and 2.
We evaluated the diagnostic accuracy and the area under the ROC curve for the two classifications. It was also assessed the agreement between the visual classification with the CAD system. As a gold-standard reference, histological results of biopsies were adopted.

Results:

The AUC for the radiologist was 0.714 and the CAD system was 0.807. The interobserver agreement according to the Kappa test was 0.8 if positive or negative results were considered. For all BI-RADS final category, including categories 2, 3, 4, and 5, the agreement was 0.58. The system impacts more at BI-RADS category 3 where a downgrade to category 2 was observed. Half of the lesions classified as BI-RADS were downgraded to 2.

Conclusion:

A CAD system can be used to classify breast masses by ultrasound with similar accuracy to a radiologist.

Limitations:

Experimental software in a small sample size.

Ethics committee approval

Approved by Institutional Ethical Committee and Brazilian Research Platform.

Funding:

No funding was received for this work.