Article Text

Download PDFPDF
1311 Adnexal masses and risk of malignancy by radiomics, external validation of a decision support tool
  1. Valentina Chiappa1,
  2. Matteo Interlenghi2,
  3. Ludovica Spanò Bascio3,
  4. Robert Fruscio4,
  5. Christian Salvatore2,
  6. Simone Ferrero5,
  7. Federica Rosati6,
  8. Silvia Ficarelli7,
  9. Lucia De Meis8,
  10. Martino Rolla9,
  11. Ida Pino10,
  12. Dorella Franchi10,
  13. Elisa Mor11,
  14. Umberto Leone Roberti Maggiore1,
  15. Giorgio Bogani1,
  16. Francesco Raspagliesi1 and
  17. Isabella Castiglioni12
  1. 1Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy
  2. 2DeepTrace Technologies S.R.L, Milan, Italy
  3. 3Ospedale Morgagni Pierantoni, Forlí, Italy
  4. 4Universitâ degli Studi di Milano-Bicocca, Monza, Italy
  5. 5San Martino Hospital and University of Genoa, Genova, Italy
  6. 6Ospedale Infermi, Rimini, Italy
  7. 7Azienda Socio Sanitaria Territoriale degli Spedali Civili di Brescia, Brescia, Italy
  8. 8IRCCS Azienda Ospedaliero-Universitaria di Bologna, Bologna, Italy
  9. 9AOU Parma,, Parma, Italy
  10. 10Istituto Europeo di Oncologia, Milan, Italy
  11. 11Ospedale Poliambulanza Brescia, Brescia, Italy
  12. 12Universitâ degli Studi di Milano-Bicocca, Milan, Italy


Introduction/Background Prospective multicenter validation of a decision support tool (TRACE 4 OC and the implemented version TRACE 4 OC+) for the triage of adnexal masses at ultrasound (US) based on radiomics and machine learning and comparison with IOTA ADNEX model.

Methodology From a multicenter consecutive series of women scheduled for surgery, we collected and evaluated 1000 US images of adnexal masses with TRACE4OC model.

All the masses were divided into 3 homogeneous groups: solid, cystic and motley and tested with TRACE4OC platform; the variables to proceed with the analysis were CA125, menopasual status, presence of shadows at US.

We then improved the previous model (TRACE4OC+) to better stratify cystic masses with ground glass echogenicity suspicious for endometriomas, often defined as medium-high risk with TRACE4OC.

The risk of malignancy of the masses was calculated also with ADNEX IOTA score (considering 10% cut off for the risk of malignancy). The performances for both the diagnostic strategies were evaluated.

Results TRACE4OC model showed 97.4% sensitivity (Se) and 79.7% specificity (Sp) when tested on the retrospective and prospective multicentric external datasets of masses (425 malignant and 575 benign lesions at final histology), achieving 77.9% PPV, 97.6% NPV and 87.7% accuracy . TRACE4OC+ model showed 96.7% Se, 83.6%Sp, 81% PPV, 97.2%NPV, 89.7% accuracy. ADNEX model -10% showed 96.7% Se, 78.4% and 76.7% PPV. Performance between the 3 models did not result in statistically significant differences (p-value>0.05).

Conclusion Radiomics and machine learning can play a promising role as a decision support tool for the clinician in the diagnostic workup of adnexal masses thus providing a reduction of the surgery rate for benign lesions still warranting very high sensitivity. Assessment with TRACE4OC and TRACE4OC+ were not shown to be inferior to IOTA ADNEX model, however requiring less information and less examiner training, representing an effective tool for the triage of adnexal masses.

Disclosures COI have been stated for all co-authors (see file uploaded).

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.