Placenta previa (PP) and Placenta Accreta Spectrum (PAS) are obstetric pathologies whose early detection is fundamental for an appropriate patient management. In this paper, ultrasonography (US) is performed on 53 patients and from the images a texture analysis feature extraction is performed through PyRadiomics. The US images were acquired with 3 different resampling resolutions: 1 × 1, 2 × 2 and 3 × 3. The features extracted from the images at each resolution were used to investigate which one is the best to make the correct diagnosis by employing machine learning techniques. Knime analytics platform was employed to implement decision tree, k nearest neighbor and naïve Bayes. Synthetic minority oversampling technique was used to balance the dataset and some evaluation metrics were computed after a leave one out cross-validation. Averaging all the metrics among all the algorithms, 1 × 1 resolution achieved the best mean accuracy (75.97%), sensitivity (83.33%), specificity (68.50%) and Area Under the Curve Receiver Operating Characteristics (0.81). Moreover, k nearest neighbor was the algorithm with the highest metrics (greater than 80%). Despite using also artificial data to balance the dataset (less than 30% of total analysed sample), this study provides researchers with the idea that employing a 1 × 1 resolution could be the best option when analysing images with machine learning algorithms on texture analysis features US-derived.

Resolution Resampling of Ultrasound Images in Placenta Previa Patients: Influence on Radiomics Data Reliability and Usefulness for Machine Learning

Cuocolo R.;
2021-01-01

Abstract

Placenta previa (PP) and Placenta Accreta Spectrum (PAS) are obstetric pathologies whose early detection is fundamental for an appropriate patient management. In this paper, ultrasonography (US) is performed on 53 patients and from the images a texture analysis feature extraction is performed through PyRadiomics. The US images were acquired with 3 different resampling resolutions: 1 × 1, 2 × 2 and 3 × 3. The features extracted from the images at each resolution were used to investigate which one is the best to make the correct diagnosis by employing machine learning techniques. Knime analytics platform was employed to implement decision tree, k nearest neighbor and naïve Bayes. Synthetic minority oversampling technique was used to balance the dataset and some evaluation metrics were computed after a leave one out cross-validation. Averaging all the metrics among all the algorithms, 1 × 1 resolution achieved the best mean accuracy (75.97%), sensitivity (83.33%), specificity (68.50%) and Area Under the Curve Receiver Operating Characteristics (0.81). Moreover, k nearest neighbor was the algorithm with the highest metrics (greater than 80%). Despite using also artificial data to balance the dataset (less than 30% of total analysed sample), this study provides researchers with the idea that employing a 1 × 1 resolution could be the best option when analysing images with machine learning algorithms on texture analysis features US-derived.
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11386/4780784
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 7
  • ???jsp.display-item.citation.isi??? ND
social impact