Medizinische Universität Graz - Research portal
Selected Publication:
Piber, R.
KI-gestützte Segmentierung abdomineller Fettkompartimente
Humanmedizin; [ Diplomarbeit ] Medizinische Universität Graz; 2022. pp. 88
[OPEN ACCESS]
FullText
- Authors Med Uni Graz:
- Advisor:
- Hassler Eva Maria
- Reishofer Gernot
- Altmetrics:
- Abstract:
- The use of artificial intelligence (AI) in radiology becomes ever more important as a way to speed up analysis and interpretation of the growing amounts of data generated by constantly improving and more widely used imaging technology. AI can be applied successfully in areas like this one, where mathematically formulating a correlation between ingoing and outgoing data proofs difficult. In the light of an already high and ever-growing number of overweight and obese people and obesity-related disorders, more efficient and accurate methods for measuring body composition – potentially AI driven - are needed. 115 young, mostly well-trained subjects were examined via the Dixon MRI technique and the data was subsequently fed into two fully automatic segmentation algorithms – one is deep-learning-based (FatSegNet), the other one is not (a k-means-based clustering algorithm), in order to learn the volume of the adipose tissue (AT) within the compartments. Furthermore, the same subjects were examined via ultrasound (US). The results generated by the different methods were then compared to each other and examined for sex differences concerning the distribution of AT over the visceral (VAT) and subcutaneous (SAT) compartments. The final analysis was performed on 92 subjects (48 females, 44 males, aged 23.53 +/- 2.84 years). The values for the volumes of the AT compartments measured by both algorithms correlate well, and the US measurement of SAT correlates tightly with the SAT volumes measured by both algorithms. Overall, the data acquired via this thesis is congruent with the current literature in showing that automatic segmentation of AT is highly accurate, even more so when done by a deep-learning-based algorithm. Also, US-based measurement of (SAT) is highly accurate and reliable, especially if done by experienced examiners implementing a sophisticated and highly standardized method. Lastly, our data is in line with the current understanding of sexual dimorphism in AT distribution patterns, with females having relatively more overall AAT (= abdominal AT = SAT+VAT) and more SAT, and males having relatively more VAT.