A Cascaded Deep Learning-Based Artificial Intelligence Algorithm for Automated Lesion Detection and Classification on Biparametric Prostate Magnetic Resonance Imaging

Abstract

Rationale and objectives: Prostate MRI improves detection of clinically significant prostate cancer; however, its diagnostic performance has wide variation. Artificial intelligence (AI) has the potential to assist radiologists in the detection and classification of prostatic lesions. Herein, we aimed to develop and test a cascaded deep learning detection and classification system trained on biparametric prostate MRI using PI-RADS for assisting radiologists during prostate MRI read out.

Materials and methods: T2-weighted, diffusion-weighted (ADC maps, high b value DWI) MRI scans obtained at 3 Tesla from two institutions (n = 1043 in-house and n = 347 Prostate-X, respectively) acquired between 2015 to 2019 were used for model training, validation, testing. All scans were retrospectively reevaluated by one radiologist. Suspicious lesions were contoured and assigned a PI-RADS category. A 3D U-Net-based deep neural network was used to train an algorithm for automated detection and segmentation of prostate MRI lesions. Two 3D residual neural network were used for a 4-class classification task to predict PI-RADS categories 2 to 5 and BPH. Training and validation used 89% (n = 1290 scans) of the data using 5 fold cross-validation, the remaining 11% (n = 150 scans) were used for independent testing. Algorithm performance at lesion level was assessed using sensitivities, positive predictive values (PPV), false discovery rates (FDR), classification accuracy, Dice similarity coefficient (DSC). Additional analysis was conducted to compare AI algorithm's lesion detection performance with targeted biopsy results.

Results: Median age was 66 years (IQR = 60-71), PSA 6.7 ng/ml (IQR = 4.7-9.9) from in-house cohort. In the independent test set, algorithm correctly detected 111 of 198 lesions leading to 56.1% (49.3%-62.6%) sensitivity. PPV was 62.7% (95% CI 54.7%-70.7%) with FDR of 37.3% (95% CI 29.3%-45.3%). Of 79 true positive lesions, 82.3% were tumor positive at targeted biopsy, whereas of 57 false negative lesions, 50.9% were benign at targeted biopsy. Median DSC for lesion segmentation was 0.359. Overall PI-RADS classification accuracy was 30.8% (95% CI 24.6%-37.8%).

Conclusion: Our cascaded U-Net, residual network architecture can detect, classify cancer suspicious lesions at prostate MRI with good detection, reasonable classification performance metrics.

Keywords: Prostate cancer; artificial intelligence; biparametric; detection; magnetic resonance imaging.

Fig 1.

Study flowchart for dataset creation for training, validation and testing of our algorithms. MRI, Magnetic Resonance Imaging.

Fig 2.

FROC curves for detection performance with BPH filter in all five folds in the validation set.

Fig 3.

FROC curves for detection performance with BPH filter in the test set.

Fig 4.

Ground truth masks and segmentation predictions in a test set patient. Green contours represent ground truth lesions and red contours are model predictions. Patient is a 66 year-old man with a PSA of 7.88 ng/ml and prostate volume of 42 ml and first-time multiparametric prostate MRI. The patient was scanned at 3T using an endorectal coil. Two distinct lesions were detected by the radiologist representing the ground truth. Lesion 1 was 0.9 cm in the right apical-mid peripheral zone and assigned PI-RADS category 4. Lesion 2 was 0.9 cm and in the left apical-mid peripheral zone and assigned PI-RADS category 4. Both lesions were correctly detected by the cascaded algorithm representing true positive lesions. (a) Original T2 weighted imaging, (b) Original Apparent Diffusion Map, (c) Original high b-value, (d) T2 weighted imaging with ground truth contours overlaid, (e) Apparent Diffusion Map with ground truth contours overlaid, (f) High b-value with ground truth contours overlaid, (g) T2 weighted imaging with prediction contours overlaid, (h) Apparent Diffusion Map with prediction contours overlaid, (i) High b-value with prediction contours overlaid.

Fig 5.

Segmentation prediction in a test set patient without associated ground truth mask. Red contours are model predictions. 71 year-old man with PSA of 6.56 ng/ml and prostate volume 51 ml. The patient was scanned at 3T using without an endorectal coil. No distinct lesions were detected representing a negative MRI (PI-RADS category 1). One false positive lesion was called by the cascaded algorithm in the right mid transition zone. The false positive lesion was reevaluated and defined as a BPH nodule. (a) Original T2 weighted imaging, (b) Original Apparent Diffusion Map, (c) Original high b-value, (d) T2 weighted imaging with prediction contours overlaid, (e) Apparent Diffusion Map with prediction contours overlaid, (f) High b-value with prediction contours overlaid.

Fig 6.

Ground truth masks and segmentation predictions in a test set patient. Green contours represent ground truth lesions and red contours are model predictions. 70 year-old man with PSA of 8.71 ng/ml and prostate volume of 63.6 ml and first-time multiparametric prostate MRI. The patient was scanned at 3T using an endorectal coil. Two distinct lesions were detected by the radiologist representing the ground truth. Lesion 1 was 1.8 cm in the midline to right apical anterior transition zone and assigned PI-RADS category 5. The second lesion was 1.6 cm in the left mid peripheral zone and assigned PI-RADS category 4. Lesion 1 was correctly detected while lesion 2 was missed by the cascaded algorithm representing a false negative. (a) Original T2 weighted imaging, (b) Original Apparent Diffusion Map, (c) Original high b-value, (d) T2 weighted imaging with ground truth contours overlaid, (e) Apparent Diffusion Map with ground truth contours overlaid, (f) High b-value with ground truth contours overlaid, (g) T2 weighted imaging with prediction contours overlaid, (h) Apparent Diffusion Map with prediction contours overlaid, (i) High b-value with prediction contours overlaid.