Artificial intelligence in the diagnosis of Parkinson’s disease from ioflupane-123 single-photon emission computed tomography dopamine transporter scans using transfer learning

Kim, Daniel H; Wit, Huub; Thurston, Mark

doi:10.1097/mnm.0000000000000890

Cited by 38 publications

(27 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…More complex methods including convolutional neural networks have been proposed for automatic classification of FP-CIT SPECT [18,20,21]. However, conventional SBR analysis is still widely used because it is easy to understand (no black box) and achieves high accuracy provided that an appropriate normal database is used.…”

Section: Introductionmentioning

confidence: 99%

Impact of the size of the normal database on the performance of the specific binding ratio in dopamine transporter SPECT

et al. 2020

View full text Add to dashboard Cite

Background: This study investigated the impact of the size of the normal database on the classification performance of the specific binding ratio (SBR) in dopamine transporter (DAT) SPECT with [ 123 I]FP-CIT in different settings. Methods: The first subject sample comprised 645 subjects from the Parkinson's Progression Marker Initiative (PPMI), 207 healthy controls (HC), and 438 Parkinson's disease (PD) patients. The second sample comprised 372 patients from clinical routine patient care, 186 with non-neurodegenerative parkinsonian syndrome (PS) and 186 with neurodegenerative PS. Single-photon emission computed tomography (SPECT) images of the clinical sample were reconstructed with two different reconstruction algorithms (filtered backprojection, iterative ordered subsets expectation maximization (OSEM) reconstruction with resolution recovery). The putaminal specific binding ratio (SBR) was computed using an anatomical region of interest (ROI) predefined in standard (MNI) space in the Automated Anatomic Labeling (AAL) atlas or using hottest voxels (HV) analysis in large predefined ROIs. SBR values were transformed to z-scores using mean and standard deviation of the SBR in a normal database of varying sizes (n = 5, 10, 15,…, 50) randomly selected from the HC subjects (PPMI sample) or the patients with non-neurodegenerative PS (clinical sample). Accuracy, sensitivity, and specificity for identifying patients with PD or neurodegenerative PS were determined as performance measures using a predefined fixed cutoff on the z-score. This was repeated for 10,000 randomly selected normal databases, separately for each size of the normal database. Mean and 5th percentile of the performance measures over the 10,000 realizations were computed. Accuracy, sensitivity, and specificity when using the whole set of HC or non-neurodegenerative PS subjects as normal database were used as benchmark.

show abstract

Section: Introductionmentioning

confidence: 99%

Impact of the size of the normal database on the performance of the specific binding ratio in dopamine transporter SPECT

et al. 2020

View full text Add to dashboard Cite

show abstract

“…Ding et al were able to show that a CNN with InceptionV3 architecture (38) could make an Alzheimer disease diagnosis with 82% specificity at 100% sensitivity (AUC, 0.98) on average 75.8 mo before the final diagnosis based on 18 F-FDG PET/CT scans and outperformed human interpreters (majority diagnosis of 5 interpreters) (39). A similar network architecture was used by Kim et al in the diagnosis of Parkinson disease from 123 I-ioflupane SPECT scans; the test sensitivity was 96.3% at 66.7% specificity (AUC, 0.87) (40). Li et al used a 3-step process of automatic segmentation, feature extraction, and classification using support vector machines and random forests to automatically detect pancreas carcinomas on 18 F-FDG PET/CT scans (41).…”

Section: Discussionmentioning

confidence: 93%

Artificial Intelligence in Nuclear Medicine

2019

View full text Add to dashboard Cite

Despite the great media attention for artificial intelligence (AI), for many health care professionals the term and the functioning of AI remain a "black box," leading to exaggerated expectations on the one hand and unfounded fears on the other. In this review, we provide a conceptual classification and a brief summary of the technical fundamentals of AI. Possible applications are discussed on the basis of a typical work flow in medical imaging, grouped by planning, scanning, interpretation, and reporting. The main limitations of current AI techniques, such as issues with interpretability or the need for large amounts of annotated data, are briefly addressed. Finally, we highlight the possible impact of AI on the nuclear medicine profession, the associated challenges and, last but not least, the opportunities.

show abstract

“…Dabei wird ein bereits für eine andere, aber verwandte Aufgabe trainiertes Lernverfahren genutzt und an die neue Fragestellung angepasst. So starteten Kim, Wit und Thurston mit einem für die Klassifikation von Fotos im Rahmen einer Large Scale Visual Recognition Challenge optimierten CNN und trainierten ausschließlich die letzte, für die finale Klassifikation zuständige Schicht des Netzes neu, und zwar für die Klassifikation von DAT-SPECTs[24]. Dazu wurden nur 108 DAT-SPECTs verwendet, davon 54 mit normaler und 54 mit reduzierter striataler DAT-Dichte, die mittels einfacher Methoden der Data Augmentation 44-fach vermehrt wurden[24].…”

unclassified

“…So starteten Kim, Wit und Thurston mit einem für die Klassifikation von Fotos im Rahmen einer Large Scale Visual Recognition Challenge optimierten CNN und trainierten ausschließlich die letzte, für die finale Klassifikation zuständige Schicht des Netzes neu, und zwar für die Klassifikation von DAT-SPECTs[24]. Dazu wurden nur 108 DAT-SPECTs verwendet, davon 54 mit normaler und 54 mit reduzierter striataler DAT-Dichte, die mittels einfacher Methoden der Data Augmentation 44-fach vermehrt wurden[24]. Das resultierende Netz lieferte eine Fläche von 0,87 unter der ROC-Kurve für die Klassifikation unabhängiger Test-DAT-SPECTs[24].…”

unclassified

“…Dazu wurden nur 108 DAT-SPECTs verwendet, davon 54 mit normaler und 54 mit reduzierter striataler DAT-Dichte, die mittels einfacher Methoden der Data Augmentation 44-fach vermehrt wurden[24]. Das resultierende Netz lieferte eine Fläche von 0,87 unter der ROC-Kurve für die Klassifikation unabhängiger Test-DAT-SPECTs[24]. Für den praktischen Einsatz ist das noch nicht ausreichend, zeigt aber, dass CNN zur automatischen Klassifikation von DAT-SPECTs auch mit sehr kleinen Datensätzen sinnvoll trainiert werden können.…”

unclassified

See 1 more Smart Citation