Click on the Scientific Publications you want to read in order to know more

Presented at SBI

This was a retrospective study on 2D mammograms from 2007-2019. The goal was to help determine and identify areas of opportunities to help radiologists when using AI and also increase patient care. The results identified that MammoScreen could help reduce recalls for suspicious calcifications. Indeed, MammoScreen was able to correctly characterize 25% of biopsied calcification clusters as negative without missing a cancer. By trusting cases assessed with low suspicion by MammoScreen, unnecessary biopsies could be reduced.

Presented at ECR 2022

Many different assets and tools are available to help educate and improve a trainee’s performance. Within this study, from Emory University, it was found that when AI is utilized during training, readers improved their performance, but also felt more confident. Trainees utilized MammoScreen as a teaching file and were able to interpret the screening more effectively and efficiently. This study helps to showcase how valuable an AI can be as an educational tool and resource for mammography readers, especially new in role.

Presented at RSNA 2021

Early detection is crucial no matter which cancer is being examined. Research has proven that screenings for breast cancer play an important role in the treatment process. 3D screenings have increased the power of screenings and AI will make screenings even more efficient.

In a study presented at RSNA 2021, Therapixel presented that MammoScreen® was able to help reduce reading time on 3D screenings. In batch readings this could lead to reading an additional 700+ cases per month! This could also mean more time is available to radiologists to get through more screenings, allow them to study other cases longer, or work on other areas of need during the day.

Additionally, during the study it was found that the use of MammoScreen increased the sensitivity by two percentage points and the specificity by five percentage points. Extrapolated on a national average, using MammoScreen, radiologists could find up to 6,600 more cancers per year. Not only will more cancers be found, but it was identified that MammoScreen use could help reduce the national recall rate by three percentage points, saving patient’s time and anxiety.

Presented at SBI 2022

Early detection of cancer is proven to increase a woman’s survival rate. In a recent US study presented at SBI 2021, 16,004 screening mammograms and their prior exams were analyzed by the MammoScreen. The results concluded that by thresholding at a MammoScreen Score™ of 6 (middle point of the MammoScreen Score, which is the best compromise in term of sensitivity and specificity), 27% of the cancer cases were found one year earlier by the AI system and 21% 2 years earlier, with a minimal increase in recall rate.

Radiology: Artificial Intelligence , vol. 2, no. 6, p. e190208, Nov. 2020

Multi-reader multi-case study to demonstrate the benefits brought by MammoScreen in the breast cancer detection process. Results have shown that the average AUC across readers was 0.769 (95% confidence interval [CI]: 0.724, 0.814) without AI and 0.797 (95% CI: 0.754, 0.840) with AI. The average difference in AUC was 0.028 (95% CI: 0.002, 0.055, P = .035). Average sensitivity was increased by 0.033 when using AI support (P = .021). Reading time changed dependently to the AI-tool score. For low likelihood of malignancy (< 2.5%) the time was about the same in the first reading session and slightly decreased in the second reading session. For higher likelihood of malignancy, the reading time was on average increased with the use of AI. We concluded that the concurrent use of this AI tool improves the diagnostic performance of radiologists in the detection of breast cancer without prolonging their workflow.

Proc. SPIE 11513, 15th International Workshop on Breast Imaging (IWBI2020), 1151321 (22 May 2020)

This study was aimed to compare the performances of MammoScreen and another deep learning based software against three radiologists ona recall-based model for mammography. A set of examinations from a daily practice, with both screening and diagnostic studies, has been interpreted by the radiologists and the two AI based algorithms. The dataset has been enriched with BIRADS 4 and 5 cases in order to have a number of cancer cases sufficient to have statistically significant results. In total, 140 examinations have been included in the final dataset. Sensitivity, False positive rate (FPR), and recall rate per BI-RADS category were considered as endpoints for each of the radiologists. While both the algorithms and radiologists have a good and comparative rate of sensitivity and FPR, the test based on BI-RADS categories (i.e. the number of cancer per BI-RADS category), showed heterogeneous results, with bad performances for one of the tested software on the extremes score of BI-RADS. Conclusions of this study say that one of the analysed software cannot be used in the current clinical practice without further improvements, MammoScreen shows promising results, but other studies are needed to have a robust external validation before being used in a daily practice.

JAMA Netw Open. 2020 Mar; 3(3): e200265.

The common final paper written together by the Organizers and Participants of the Digital Mammography DREAM Challenge where Therapixel was ranked 1st. The Dream Challenge has been described as the largest data challenge to date focused on mammographic imaging. This paper evaluates the AI models submitted for mammography analysis. The ensemble of models created by the Top-8 teams were evaluated on two blind datasets from the US and Sweden, in general they performed coherently across the different datasets. While the AI ensemble still underperformed with respect to a dedicated fellowship trained mammographer, it did however demonstrate that radiologists working in conjunction with the AI assessments lead to a statistically significant improvement of sensitivity, from 90.5% (radiologist only) to 92% (radiologist with AI). In the US, this would have reduced the number of women requiring unnecessary diagnostic workup by more than half a million. Since the challenge Therapixel has significantly improved its AI models by more than doubling the amount of data used to initially develop the MammoScreen system.

Presented at RSNA 2021

One beneficial tool for radiologists is using prior screenings to compare to current screenings and observe any changes. Incorporating priors into the AI’s consideration set, can the specificity of the AI improve? MammoScreen® set out to answer that question!

In a recent study presented at RSNA 2021, from a dataset of over 50,000 screenings, with at least one prior in the range 6-18 months, and dissected to a subset of 5,848. From there, AI considered 858 studies as suspicious (MammoScreen score of 6 or higher). Of those cases, 536 were cancer positive, or 62%.

With the use of priors, for the highest suspicious scores, MammoScreen was able to almost double the amount of identified cancer cases without creating a false positive. MammoScreen utilizing priors can be used with more confidence for immediate recalls.

Poster Presented at EUSOBI 2022

An AI tool, originally developed to provide a level of suspicion independently on mammography or tomosynthesis images, showed superior performance when predictions on both modalities are merged compared to AIs taken individually on each modality. This system could potentially improve efficiency of both DBT and mammography.

Presented at SBI 2023

As for humans, AI has performance linked to lesion visibility on the used technique. Tomosynthesis (3D) is better for identifying soft tissue lesions, and calcifications are easier to spot on FFDM (2D). MammoScreen worked to identify if these modalities could be combined to work better together. To find the solution to this question, over 3,000 patients that underwent a screening mammogram were evaluated, containing 750+ cancer cases and 2300+ negative screenings. The results indicated that the combination of both modalities outperformed 2D and 3D performances when reviewed individually.