Artificial Intelligence in Medicine: Dr. Sperling’s Early Contributions to AI in Imaging

Considering ever-developing advances in computational power and availability of large data sets, the marriage of humans and machines in future clinical practice seems inevitable. Therefore, regardless of their feelings, the radiologists should be familiar with these concepts.^[i]

A seemingly unquenchable thirst for efficient, highly accurate and patient-friendly imaging drives the ongoing integration of Artificial Intelligence (AI) in radiologic imaging. Magnetic Resonance Imaging particularly lends itself to this quest. In part, this is due to the lack of exposure to ionizing radiation, a safety feature that brings this highly accurate imaging to a limitless number of patients, including repeat scans. It is also due to the physics of MRI that it benefits from engineering innovations. Thus, there is a natural marriage of AI (including machine learning and deep learning) with MRI’s biomedical engineering technology and methods.

Dr. Dan Sperling and colleagues from Case Western Reserve University (notably Anant Madabhushi, PhD) and other institutions have been leaders in this integration since the early years of the new millennium. In particular, their published research has focused on the application of this marriage to prostate imaging, and MRI-guided Focal Laser Ablation for the treatment of prostate tumors.

Improving prostate MRI with AI methods

Multiparametric MRI sets a high bar for prostate scanning. Two or more parameters (imaging sequences) each produce defining anatomic and tissue quality characteristics. When combined, they offer a detailed portrait of prostate health or disease, with information on the location, extent and risk level of abnormal lesions.

Before presenting four publications for which Sperling and Madabhushi are co-authors, here are a few summaries of how AI applications figure into their studies:

• The Finite Element Method (FEM) is a computational method adapted from mechanical engineering, used to generate models that approximate the real world. The value of FEM for prostate imaging is its ability to quantify intraprostatic changes from one state to another (e.g. pre-treatment prostate vs. post-treatment prostate). With AI, very large data sets, both mathematical and clinical, can be consulted at any step in the process.
• AI facilitates comparison at a detailed level imperceptible to human eyes. When trained for pattern recognition and textural analysis, a computer learns to extract and co-register magnified imaging voxels from one parameter to another, allowing voxel-to-voxel co-registration of visual information in generating a 3D prostate map. In addition to qualitative information from each parameter, voxel-based quantitative analysis can then utilize AI probability algorithms to predict responses to treatment.
• Conventional models of MRI-based prostate diagnostic imaging have relied on anatomic landmarks to identify tissue segments such as prostate zones (transitional, central and peripheral). Development of alternative trained networks that bypass ambiguous anatomic features can assist diagnostic tasks that rely on accurate gland segmentation and cancer detection. “Along with the prostate, the training dataset size could be varied for other abdominal organs such as the kidney.”^[ii]

Four published studies

To date, Sperling, Madabhushi and others have contributed the following studies to the body of work pertaining to prostate MRI and AI:

1. Toth R, Sperling D, Madabhushi A. Quantifying Post- Laser Ablation Prostate Therapy Changes on MRI via a Domain-Specific Biomechanical Model: Preliminary Findings.^[iii] The team generated a physical biomechanical FEM model to quantify pre- and post-Focal Laser Ablation (FLA) induced changes in the prostate morphology and structure, based on mpMRI scans. They note that their system of quantitative measurements reflecting deformation changes could be used to track treatment response over time.
2. Viswanath S, Toth R, Rusu M, Sperling D, Lepor H, Futterer J, Madabhushi A. Quantitative Evaluation of Treatment Related Changes on Multi-Parametric MRI after Laser Interstitial Thermal Therapy of Prostate Cancer.^[iv] In 2013, the authors delivered a presentation of a new voxel-by-voxel quantitative method of co-registering (aligning) T2 weighted and diffusion weighted parameters within each patient’s pre- and post-FLA images; the pair of aligned pre/post images were aligned with each other. The purpose was “to attempt to (1) quantitatively identify MP-MRI markers predictive of favorable treatment response and longer term patient outcome, and (2) identify which MP-MRI markers are most sensitive to post-[FLA] changes in the prostate.” They found that “T2w texture may be highly sensitive as well as specific” in modeling changes, suggesting that modeling non-invasive MP-MRI imaging markers could potentially help determine long- and short-term patient outcomes.
3. Viswanath S, Toth R, Rusu M, Sperling D, Lepor H, Futterer J, Madabhushi A. Identifying Quantitative In Vivo Multi-Parametric MRI Features for Treatment Related Changes after Laser Interstitial Thermal Therapy of Prostate Cancer.^[v] This 2014 published paper is a more developed and specific version of the 2013 presentation. In addition to the voxel-wise comparison of pre- and post-treatment aligned images, further quantification was gained by correcting for intensity drift in order to examine tissue-specific response, plus the information captured by T2w MRI and ADC maps via texture and intensity features. This process “resulted in visually discernible improvements in highlighting tissue-specific response in different MRI features…” indicating that instead of the original parametric images and ADC values, “…steerable and non-steerable gradient texture features…were highly sensitive as well as specific in identifying changes within the ablation zone…” This method of distilling a high number of quantitative features from MRI images opens the door for coupling with AI, which can handle “a massive amount of data compared with traditional statistical methods.”^[vi]
4. Toth R, Ribault J, Gentile J, Sperling D, Madabhushi A. Simultaneous Segmentation of Prostatic Zones Using Active Appearance Models With Multiple Coupled Levelsets.^[vii] This paper presents an improved algorithm that can be used in the training phase of deep learning for image recognition. The widely-used Active Appearance Model (AAM) algorithm defines “the shape of each object using a set of anatomical landmarks,” and, as noted early, these can be ambiguous and hard to identify clearly; also, AAMs allow for segmentation of only a single object. As an alternative, the team developed what they call Multiple-Levelset AAM (MLA). According to their paper, the MLA “can simultaneously segment multiple objects, and makes use of multiple levelsets, rather than anatomical landmarks, to define the shapes.” When applied to prostate imaging, a hierarchical segmentation framework that could utilize domain-specific characteristics in order to achieve greater accuracy in modeling prostate zonal anatomy.

When viewed as a set, these four works establish the contributions of Dr. Sperling and his cohorts to the union of radiomics (the extraction of a large number of quantitative features from medical images) with AI. Key elements of diseased tissue that would otherwise remain hidden become accessible for diagnosis and decision-making. Dr. Sperling’s grounding in AI for radiology is the foundation for his ongoing expertise and authority in this rapidly growing field.

NOTE: This content is solely for purposes of information and does not substitute for diagnostic or medical advice. Talk to your doctor if you are experiencing pelvic pain, or have any other health concerns or questions of a personal medical nature.

---

[i] Koçak B, Durmaz E, Ates E, K?l?çkesmez Ö. Radiomics with artificial intelligence: a practical guide for beginners. Diagn Interv Radiol. 2019 Nov; 25(6): 485–495.
[ii] Bardis M, Houshyar R, Chanon C et al. Deep Learning with Limited Data: Organ Segmentation Performance by U-Net. July 2020 Electronics 9(8):1199
[iii] Toth R, Sperling D, Madabhushi A. Quantifying Post- Laser Ablation Prostate Therapy Changes on MRI via a Domain-Specific Biomechanical Model: Preliminary Findings. PLoS ONE. 2016 Apr; 11(4): e0150016. DOI: 10.1371/journal.pone.0150016
[iv] Viswanath S, Toth R, Rusu M, Sperling D, Lepor H, Futterer J, Madabhushi A. Quantitative Evaluation of Treatment Related Changes on Multi-Parametric MRI after Laser Interstitial Thermal Therapy of Prostate Cancer. Proceedings – Society of Photo-Optical Instrumentation Engineers. 2013 Mar; 8671:86711F.
DOI: 10.1117/12.2008037
[v] Viswanath S, Toth R, Rusu M, Sperling D, Lepor H, Futterer J, Madabhushi A. Identifying Quantitative In Vivo Multi-Parametric MRI Features For Treatment Related Changes after Laser Interstitial Thermal Therapy of Prostate Cancer. Neurocomputing. 2014 Nov 20; 144: 13-23. DOI: 10.1016/j.neucom.2014.03.065
[vi] Koçak B, et al. Ibid.
[vii] Toth R, Ribault J, Gentile J, Sperling D, Madabhushi A. Simultaneous Segmentation of Prostatic Zones Using Active Appearance Models With Multiple Coupled Levelsets. Computer Vision and Image Understanding. 2013 Aug; 117(9):1051-1060. DOI: 10.1016/j.cviu.2012.11.013