Optimized AI-Assisted Diagnosis of Spinal Anomalies Using Convolutional Neural Networks by Enhancing Feature Extraction in Small Datasets

Abstract

Purpose: Major spinal health anomalies, particularly idiopathic scoliosis and spondylolisthesis are primarily caused by abnormal vertebral displacements. Early diagnosis is critical for effective treatment and management. However, diagnosis of these conditions requires the analysis of X-ray images by expert physicians, and when the number of patients increases, the amount of required time to have a diagnosis may take longer duration. Also, the concentration of the physician may be lost. As a result of this, physician may have an erroneous decision related to diagnosis. As a solution to the problem, we suggest a method based on artificial intelligence that helps physician to come up with the correct diagnosis. Materials and Methods: To address the issue of insufficient datasets, we use a customized convolutional neural network model and the Leaky ReLU activation function. This approach helps us extract better features while reducing computational complexity. Results: In our experiments, we achieve success rates of 98.51% in accuracy, 98.63% in precision, 98.53% in recall, and 98.51% in the F1 score. When we compare these results to another study using the same dataset, we see increases of 2.25% in accuracy, 1.04% in precision, 2.67% in recall, and 4.11% in the F1 score. To avoid misleading results from small or imbalanced datasets, we use a balanced version of the dataset for comparison. When we compare the model trained on the imbalanced dataset with the version trained on the balanced dataset, we find a minimal performance decrease of only 0.787% in the F1 score and an average decrease of 0.721% in the other metrics. This shows that the model performs well regardless of potential issues from dataset imbalance. We also test the model with challenging data and obtain successful metrics. Conclusion: We achieve the objectives of increasing the success rate by reducing computational complexity and improving feature extraction for small datasets. Furthermore, experiments with challenging datasets show that our method remains generalizable and usable even on small datasets.

Keywords

• Spine health
• Scoliosis
• Artificial intelligence
• Convolutional neural network
• Leaky ReLU

Küçük Veri Setlerinde Özellik Çıkarmayı Geliştirerek Evrişimli Sinir Ağları ile Omurga Anomalilerinin Yapay Zeka Destekli Teşhisinin Optimizasyonu

Öz

Amaç: Omurga sağlığıyla ilgili başlıca anomaliler, özellikle idiopatik skolyoz ve spondilolistezis, esasen anormal vertebral kaymalar nedeniyle ortaya çıkar. Erken teşhis, etkili tedavi ve yönetim için kritiktir. Ancak, bu durumların teşhisi, X-ray görüntülerinin uzman hekimler tarafından analiz edilmesini gerektirir ve hasta sayısı arttığında teşhis süresi daha uzun olabilir. Ayrıca, hekimin dikkati dağılabilir. Bunun sonucunda, hekim teşhisle ilgili yanlış bir karar verebilir. Soruna çözüm olarak, hekime doğru teşhis koymasına yardımcı olacak yapay zeka temelli bir yöntem öneririz. Gereç ve Yöntem: Veri seti yetersizliği sorununu ele almak için, özelleştirilmiş bir konvolüsyonel sinir ağı modeli ve Leaky ReLU aktivasyon fonksiyonunu kullanırız. Bu yaklaşım, daha iyi özellik çıkarımı yapmamıza ve hesaplama karmaşıklığını azaltmamıza yardımcı olur. Bulgular: Deneylerimizde, doğrulukta %98.51, kesinlikte %98.63, duyarlılık %98.53 ve F1 skorunda %98.51 başarı oranları elde ediyoruz. Bu sonuçları aynı veri setini kullanan başka bir çalışma ile karşılaştırdığımızda, sırasıyla doğrulukta %2.25, kesinlikte %1.04, geri çağırmada %2.67 ve F1 skorunda %4.11 artışlar görüyoruz. Küçük veya dengesiz veri setlerinden kaynaklanabilecek yanıltıcı sonuçları önlemek için, karşılaştırma için dengelenmiş bir veri seti kullanırız. Dengesiz veri seti ile eğitilen modeli, dengelenmiş veri seti ile eğitilen versiyonu ile karşılaştırdığımızda, F1 skorunda yalnızca %0.787'lik minimal bir performans düşüşü ve diğer metriklerde ortalama %0.721'lik bir düşüş buluruz. Bu, modelin veri setinin dengesizliklerinden kaynaklanabilecek potansiyel sorunlara rağmen iyi performans gösterdiğini gösterir. Ayrıca, modeli zorlu verilerle test ediyoruz ve başarılı metrikler elde ediyoruz. Sonuç: Hesaplama karmaşıklığını azaltarak ve küçük veri setleri için özellik çıkarımını artırarak başarı oranını artırma hedeflerini yakalarız. Ayrıca, zorlu veri setleri ile yapılan deneyler, yöntemimizin küçük veri setlerinde bile genellenebilir ve kullanılabilir olduğunu gösterir.

Anahtar Kelimeler

• Omurga sağlığı
• Skolyoz
• Yapay zekâ
• Evrişimli sinir ağı
• Sızdıran ReLU

References

1. 1. Rohit Aiyer. Chapter 1 - an overview on the anatomy of the spine. In Alaa Abd-Elsayed, editor, Decompressive Techniques (First Edition), Atlas of Interventional Pain Management Series, pages 1–12. Elsevier, New Delhi, first edition, 2024.
2. 2. Adrese Michael Kandahari, Varun Puvanesarajah, Francis H. Shen, Jon Raso, and Hamid Hassanzadeh. 1 - anatomy of the spine. In Dino Samartzis, Jaro I. Karppinen, and Frances M.K. Williams, editors, Spine Phenotypes, pages 1–34. Academic Press, 2022.
3. 3. John H. Bland and Dallas R. Boushey. Anatomy and physiology of the cervical spine. Seminars in Arthritis and Rheumatism, 20(1):1–20, 1990.
4. 4. Max Aebi. The adult scoliosis. European spine journal, 14:925–948, 2005.
5. 5. Norman Capener. Spondylolisthesis. The British Journal of Surgery, volume 19, pages 374-386, 1932.
6. 6. Jack C Cheng, René M Castelein, Winnie C Chu, Aina J Danielsson, Matthew B Dobbs, Theodoros B Grivas, Christina A Gurnett, Keith D Luk, Alain Moreau, Peter O Newton, et al. Adolescent idiopathic scoliosis. Nature reviews disease primers, 1(1):1–21, 2015.
7. 7. Alex MacLennan. Scoliosis. The British Medical Journal, pages 864–866, 1922.
8. 8. Caroline J Goldberg, David P Moore, Esmond E Fogarty, and Frank E Dowling. Scoliosis: a review. Pediatric surgery international, 24:129–144, 2008.

9. 9. JA Fitzgerald and PH Newman. Degenerative spondylolisthesis. The Journal of Bone & Joint Surgery British Volume, 58(2):184–192, 1976.
10. 10. Joseph A Janicki and Benjamin Alman. Scoliosis: Review of diagnosis and treatment. Paediatrics & child health, 12(9):771–776, 2007.
11. 11. Joseph S Lombardi, Leon L Wiltse, James Reynolds, Eric H Widell, and CURTIS SPENCER III. Treatment of degenerative spondylolisthesis. Spine, 10(9):821–827, 1985.
12. 12. JOHN R FISK, JOHN H MOE, and ROBERT B WINTER. Scoliosis,
13. 13. spondylolysis, and spondylolisthesis: their relationship as reviewed in 539 patients. Spine, 3(3):234–245, 1978.
14. 14. Yeliz Basar, Deniz Alis, Deniz Esin Tekcan Sanli, Tugana Akbas, and Ercan Karaarslan. Whole-body mri for preventive health screening: Management strategies and clinical implications. European Journal of Radiology, 137:109584, 2021.
15. 15. Tomasz Kotwicki. Evaluation of scoliosis today: examination, x-rays and beyond. Disability and rehabilitation, 30(10):742–751, 2008.
16. 16. Pratik Shrestha, Aachal Singh, Riya Garg, Ishika Sarraf, TR Mahesh, and G Sindhu Madhuri. Early stage detection of scoliosis using machine learning algorithms. In 2021 International Conference on Forensics, Analytics, Big Data, Security (FABS), volume 1, pages 1–4. IEEE, 2021.
17. 17. Peter Bernstein, Johannes Metzler, Marlene Weinzierl, Carl Seifert, Wadim Kisel, and Markus Wacker. Radiographic scoliosis angle esti- mation: spline-based measurement reveals superior reliability compared to traditional cobb method. European spine journal, 30:676–685, 2021.
18. 18. Joddat Fatima, Mashood Mohsan, Amina Jameel, Muhammad Usman Akram, and Adeel Muzaffar Syed. Vertebrae localization and spine segmentation on radiographic images for featurebased curvature clas- sification for scoliosis. Concurrency and Computation: Practice and Experience, 34(26):e7300, 2022.
19. 19. Kai Chen, Xiao Zhai, Kaiqiang Sun, Haojue Wang, Changwei Yang, and Ming Li. A narrative review of machine learning as promising revolution in clinical practice of scoliosis. Annals of Translational Medicine, 9(1), 2021.
20. 20. Torgyn Shaikhina and Natalia A Khovanova. Handling limited datasets with neural networks in medical applications: A small-data approach. Artificial intelligence in medicine, 75:51–63, 2017.
21. 21. Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3. Atlanta, GA, 2013.
22. 22. Andrinandrasana David Rasamoelina, Fouzia Adjailia, and Peter Sincák. A review of activation function for artificial neural network. In 2020 IEEE 18th World Symposium on Applied Machine Intelligence and Informatics (SAMI), pages 281–286. IEEE, 2020.
23. 23. Bin Ding, Huimin Qian, and Jun Zhou. Activation functions and their characteristics in deep neural networks. In 2018 Chinese control and decision conference (CCDC), pages 1836–1841. IEEE, 2018.
24. 24. Sepp Hochreiter. The vanishing gradient problem during learning recurrent neural nets and problem solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 6(02):107–116, 1998.
25. 25. Lu Lu, Yeonjong Shin, Yanhui Su, and George Em Karniadakis. Dying relu and initialization: Theory and numerical examples. arXiv preprint arXiv:1903.06733, 2019.
26. 26. Scott C Douglas and Jiutian Yu. Why relu units sometimes die: Analysis of single-unit error backpropagation in neural networks. In 2018 52nd Asilomar Conference on Signals, Systems, and Computers, pages 864– 868. IEEE, 2018.
27. 27. Mohammad Fraiwan, Ziad Audat, Luay Fraiwan, and Tarek Manasreh. Using deep transfer learning to detect scoliosis and spondylolisthesis from x-ray images. Plos one, 17(5):e0267851, 2022.
28. 28. Hieu T Nguyen, Hieu H Pham, Nghia T Nguyen, Ha Q Nguyen, Thang Q Huynh, Minh Dao, and Van Vu. Vindr-spinexr: A deep learning framework for spinal lesions detection and classification from radiographs. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, 9 September 27–October 1, 2021, Proceedings, Part V 24, pages 291–301. Springer, 2021.
29. 29. Scoliosis test dataset, accurate automated spinal curvature estimation. Online, 2019.
30. 30. Eli Stevens, Luca Antiga, and Thomas Viehmann. Deep learning with PyTorch. Manning Publications, 2020.
31. 31. Ekaba Bisong and Ekaba Bisong. Google colaboratory. Building machine learning and deep learning models on google cloud platform: a comprehensive guide for beginners, pages 59–64, 2019.
32. 32. Alexandra L’heureux, Katarina Grolinger, Hany F Elyamany, and Mi- riam AM Capretz. Machine learning with big data: Challenges and approaches. Ieee Access, 5:7776–7797, 2017.
33. 33. Lina Zhou, Shimei Pan, Jianwu Wang, and Athanasios V Vasilakos. Machine learning on big data: Opportunities and challenges. Neurocomputing, 237:350–361, 2017.