An objective assessment of the differences between synthetic and digital mammography using AI
(2026) MSFT02 20262Medical Physics Programme
- Abstract
- Background and purpose: Breast cancer is the most common type of cancer among women in Sweden. The mortality can be significantly reduced if the breast cancer is detected at an early stage. Screening programs in Sweden are offered for all woman between 40-74 years, using conventional digital mammography (DM). Digital breast tomosynthesis (DBT) is superior to DM in terms of diagnostic accuracy, but has a higher work- load than DM. From the DBT image volume, a synthetic mammography (SM) image is reconstructed. A potential screening strategy would be to image all women with DBT and let an AI system triage the women depending on their breast cancer risk to be investigated by SM (low-risk women, which is a majority of the women in a screening... (More)
- Background and purpose: Breast cancer is the most common type of cancer among women in Sweden. The mortality can be significantly reduced if the breast cancer is detected at an early stage. Screening programs in Sweden are offered for all woman between 40-74 years, using conventional digital mammography (DM). Digital breast tomosynthesis (DBT) is superior to DM in terms of diagnostic accuracy, but has a higher work- load than DM. From the DBT image volume, a synthetic mammography (SM) image is reconstructed. A potential screening strategy would be to image all women with DBT and let an AI system triage the women depending on their breast cancer risk to be investigated by SM (low-risk women, which is a majority of the women in a screening population), or their DBT images (high-risk women).
The SM images could potentially be used as a substitute for DM, to achieve a suitable combination of SM for low-risk women, and DBT for high-risk women. This study aims to investigate whether SM reconstructed from DBT image volumes is equivalent to DM regarding diagnostic accuracy, by using AI cancer detection software.
Material and Methods: One hundred women recalled from screening were invited to participate in this preparatory study. The women who accepted to participate underwent imaging using two different modalities, DBT and DM. SM was reconstructed from the DBT volumes. The imaging is performed at Unilabs, Malmö, Sweden, using a clinically used mammography screening device (B.brilliant, Siemens Healthineers, Forchheim, Germany). An AI system, Transpara (Screenpoint Medical, Nijmegen, NL) was used to analyze the images. For this study to compare SM to DM, the AI-system needed to explicitly process only SM. To accomplish this, the SM images were presented in the same format as DM images, by changing the DICOM-information and the image background. The change of the image backgrounds was necessary as DBT and SM images were cropped, unlike DM which used a set field of view. The AI-system provided a risk score that represents the level of cancer suspicion. The AI risk scores of the SM and DM with the malignant cases was compared using a statistical test to test the null hypothesis that there is no significant difference between the modalities. Further analysis of all 100 cases was done using ROC analysis to evaluate diagnostic accuracy.
Results: Among the malignant cases, the SM yielded consistently higher risk score than DM. A Wilcoxon signed-rank test confirmed a statistically significant paired difference between modalities (p=3.597×10⁻⁶). From the ROC analysis of the 100 cases, the resulting area under the curve, AUC, for DM was 0.843 compared to the AUC for SM which was 0.753. Based on DeLong’s test the difference in AUC indicated that there was no significant difference between the two modalities (p=0.071).
Conclusion: These results indicate that SM provides AI risk scores comparable to, or higher, than DM for malignant cases. The ROC curve for SM showed a tendency toward higher false-positive rates, suggesting increased sensitivity. If further studies confirm that the diagnostic accuracy between the modalities show no significant difference, then these findings of SM are supporting its suitability as a potential substitute for DM images in future screening programs. (Less) - Popular Abstract (Swedish)
- När kvinnor genomgår en mammografiundersökning så tas olika typer av röntgenbilder för att
kunna upptäcka bröstcancer. Regelbundna mammografiundersökningar är viktiga eftersom
tidig upptäckt kraftigt minskar dödligheten från bröstcancer. Nya tekniker utvecklas konstant
för att förbättra hur väl man kan undersöka bröstet utan att göra en invasiv undersökning,
samtidigt som inte vill öka stråldosen. En typ av sådan teknik är den syntetiska
mammografibilden (SM) som skapas utifrån en 3D-undersökning. Denna 3D-undersökning
kallas digital brösttomosyntes (DBT). En SM-bild produceras i stället för att det skall tas en ny
vanlig 2D-bild utöver undersökningen.
I detta arbete så undersöker jag hur väl dessa SM-bilder fungerar när de... (More) - När kvinnor genomgår en mammografiundersökning så tas olika typer av röntgenbilder för att
kunna upptäcka bröstcancer. Regelbundna mammografiundersökningar är viktiga eftersom
tidig upptäckt kraftigt minskar dödligheten från bröstcancer. Nya tekniker utvecklas konstant
för att förbättra hur väl man kan undersöka bröstet utan att göra en invasiv undersökning,
samtidigt som inte vill öka stråldosen. En typ av sådan teknik är den syntetiska
mammografibilden (SM) som skapas utifrån en 3D-undersökning. Denna 3D-undersökning
kallas digital brösttomosyntes (DBT). En SM-bild produceras i stället för att det skall tas en ny
vanlig 2D-bild utöver undersökningen.
I detta arbete så undersöker jag hur väl dessa SM-bilder fungerar när de analyseras av ett AI-
system, och hur de presterar jämfört med de traditionella 2D-bilderna (DM) som används i
dagens mammografiundersökning. Eftersom SM-bilderna inte har exakt samma format som
DM, behövde de först förberedas så att AI-systemet kunde tolka dem likvärdigt. När detta var
gjort fick AI-systemet analysera båda bildtyperna, och resultaten jämfördes för att se om SM
kan fungera som ett tillförlitligt alternativ till DM.
Analysen visar hur väl detta AI-system kan arbeta med syntetiska bilder, men att SM tenderar
att vara mer försiktig i sina bedömningar. I denna studie innebär det att fler kvinnor klassades
som misstänkta för bröstcancer trots att de inte hade någon konstaterad sjukdom. Det betyder
inte att SM är en dålig metod, men det visar att tekniken behöver utvärderas noggrant innan
den kan ersätta DM i klinisk rutin. En alltför försiktig metod kan leda till fler återkallade, vilket
kan skapa oro för patienterna och belasta vården. Samtidigt har SM stora fördelar. Eftersom
bilden skapas från DBT-undersökningen får radiologerna både 2D- och 3D-information från
samma undersökningstillfälle, utan att stråldosen ökar. Detta gör tekniken attraktiv för
framtidens mammografi, där målet är att kombinera låg stråldos med hög detaljrikedom.
För att kunna dra säkra slutsatser behövs dock större studier. Den här studien visar tydligt att
SM har potential, men också att det finns utmaningar som måste hanteras innan tekniken kan
användas fullt ut. Framför allt behövs det mer forskning om hur AI-systemet ska tränas och
anpassas för att tolka SM-bilder på ett sätt som motsvarar eller överträffar dagens DM-
baserade metoder. Med större datamängder och fler deltagare kan man få mer heltäckande
bild av hur SM fungerar i praktiken.
Sammanfattningsvis så visar projektet att SM är en lovande teknik som kan bidra till mer
skonsamma och effektiva mammografiundersökningar i framtiden. Men för att tekniken ska
kunna införas brett krävs fortsatt forskning, noggranna jämförelser och större studier. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9248500
- author
- Kristoffersson, Emma
- supervisor
- organization
- course
- MSFT02 20262
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- language
- English
- id
- 9248500
- date added to LUP
- 2026-08-17 16:29:11
- date last changed
- 2026-08-18 11:15:31
@misc{9248500,
abstract = {{Background and purpose: Breast cancer is the most common type of cancer among women in Sweden. The mortality can be significantly reduced if the breast cancer is detected at an early stage. Screening programs in Sweden are offered for all woman between 40-74 years, using conventional digital mammography (DM). Digital breast tomosynthesis (DBT) is superior to DM in terms of diagnostic accuracy, but has a higher work- load than DM. From the DBT image volume, a synthetic mammography (SM) image is reconstructed. A potential screening strategy would be to image all women with DBT and let an AI system triage the women depending on their breast cancer risk to be investigated by SM (low-risk women, which is a majority of the women in a screening population), or their DBT images (high-risk women).
The SM images could potentially be used as a substitute for DM, to achieve a suitable combination of SM for low-risk women, and DBT for high-risk women. This study aims to investigate whether SM reconstructed from DBT image volumes is equivalent to DM regarding diagnostic accuracy, by using AI cancer detection software.
Material and Methods: One hundred women recalled from screening were invited to participate in this preparatory study. The women who accepted to participate underwent imaging using two different modalities, DBT and DM. SM was reconstructed from the DBT volumes. The imaging is performed at Unilabs, Malmö, Sweden, using a clinically used mammography screening device (B.brilliant, Siemens Healthineers, Forchheim, Germany). An AI system, Transpara (Screenpoint Medical, Nijmegen, NL) was used to analyze the images. For this study to compare SM to DM, the AI-system needed to explicitly process only SM. To accomplish this, the SM images were presented in the same format as DM images, by changing the DICOM-information and the image background. The change of the image backgrounds was necessary as DBT and SM images were cropped, unlike DM which used a set field of view. The AI-system provided a risk score that represents the level of cancer suspicion. The AI risk scores of the SM and DM with the malignant cases was compared using a statistical test to test the null hypothesis that there is no significant difference between the modalities. Further analysis of all 100 cases was done using ROC analysis to evaluate diagnostic accuracy.
Results: Among the malignant cases, the SM yielded consistently higher risk score than DM. A Wilcoxon signed-rank test confirmed a statistically significant paired difference between modalities (p=3.597×10⁻⁶). From the ROC analysis of the 100 cases, the resulting area under the curve, AUC, for DM was 0.843 compared to the AUC for SM which was 0.753. Based on DeLong’s test the difference in AUC indicated that there was no significant difference between the two modalities (p=0.071).
Conclusion: These results indicate that SM provides AI risk scores comparable to, or higher, than DM for malignant cases. The ROC curve for SM showed a tendency toward higher false-positive rates, suggesting increased sensitivity. If further studies confirm that the diagnostic accuracy between the modalities show no significant difference, then these findings of SM are supporting its suitability as a potential substitute for DM images in future screening programs.}},
author = {{Kristoffersson, Emma}},
language = {{eng}},
note = {{Student Paper}},
title = {{An objective assessment of the differences between synthetic and digital mammography using AI}},
year = {{2026}},
}