Automatic segmentation of human knee anatomy by a convolutional neural network applying a 3D MRI protocol.

BACKGROUND: To study deep learning segmentation of knee anatomy with 13 anatomical classes by using a magnetic resonance (MR) protocol of four three-dimensional (3D) pulse sequences, and evaluate possible clinical usefulness. METHODS: The sample selection involved 40 healthy right knee volumes from adult participants. Further, a recently injured single left knee with previous known ACL reconstruction was included as a test subject. The MR protocol consisted of the following 3D pulse sequences: T1 TSE, PD TSE, PD FS TSE, and Angio GE. The DenseVNet neural network was considered for these experiments. Five input combinations of sequences (i) T1, (ii) T1 and FS, (iii) PD and FS, (iv) T1, PD, and FS and (v) T1, PD, FS and Angio were trained using the deep learning algorithm. The Dice similarity coefficient (DSC), Jaccard index and Hausdorff were used to compare the performance of the networks. RESULTS: Combining all sequences collectively performed significantly better than other alternatives. The following DSCs (±standard deviation) were obtained for the test dataset: Bone medulla 0.997 (±0.002), PCL 0.973 (±0.015), ACL 0.964 (±0.022), muscle 0.998 (±0.001), cartilage 0.966 (±0.018), bone cortex 0.980 (±0.010), arteries 0.943 (±0.038), collateral ligaments 0.919 (± 0.069), tendons 0.982 (±0.005), meniscus 0.955 (±0.032), adipose tissue 0.998 (±0.001), veins 0.980 (±0.010) and nerves 0.921 (±0.071). The deep learning network correctly identified the anterior cruciate ligament (ACL) tear of the left knee, thus indicating a future aid to orthopaedics. CONCLUSIONS: The convolutional neural network proves highly capable of correctly labeling all anatomical structures of the knee joint when applied to 3D MR sequences. We have demonstrated that this deep learning model is capable of automatized segmentation that may give 3D models and discover pathology. Both useful for a preoperative evaluation.

Placental size at gestational week 27 and 37: The associations with pulsatility index in the uterine and the fetal-placental arteries.

INTRODUCTION: Fetal growth restriction is known to be related to decreased fetal and placental blood flow. It is not known, however, whether placental size is related to fetal and placental blood flow. We studied the correlations of intrauterine placental volume and placental-fetal-ratio with pulsatility index (PI) in the uterine arteries, fetal middle cerebral artery, and umbilical artery. METHODS: We followed a convenience sample of 104 singleton pregnancies, and we measured placental and fetal volumes using magnetic resonance imaging (MRI) at gestational week 27 and 37 (n = 89). Pulsatility index (PI) was measured using Doppler ultrasound. We calculated cerebroplacental ratio as fetal middle cerebral artery PI/umbilical artery PI and placental-fetal-ratio as placental volume (cm3)/fetal volume (cm3). RESULTS: At gestational week 27, placental volume was negatively correlated with uterine artery PI (r = -0.237, p = 0.015, Pearson's correlation coefficient), and positively correlated with fetal middle cerebral artery PI (r = 0.247, p = 0.012) and cerebroplacental ratio (r = 0.208, p = 0.035). Corresponding correlations for placental-fetal-ratio were -0.273 (p = 0.005), 0.233 (p = 0.018) and 0.183 (p = 0.064). Umbilical artery PI was not correlated with placental volume. At gestational week 37, we found weaker and no significant correlations between placental volume and the pulsatility indices. CONCLUSIONS: Our results suggest that placental size is correlated with placental and fetal blood flow at gestational week 27.

Automatic placental and fetal volume estimation by a convolutional neural network.

INTRODUCTION: We aimed to develop an artificial intelligence (AI) deep learning algorithm to efficiently estimate placental and fetal volumes from magnetic resonance (MR) scans. METHODS: Manually annotated images from an MRI sequence was used as input to the neural network DenseVNet. We included data from 193 normal pregnancies at gestational week 27 and 37. The data were split into 163 scans for training, 10 scans for validation and 20 scans for testing. The neural network segmentations were compared to the manual annotation (ground truth) using the Dice Score Coefficient (DSC). RESULTS: The mean ground truth placental volume at gestational week 27 and 37 was 571 cm3 (Standard Deviation (SD) 293 cm3) and 853 cm3 (SD 186 cm3), respectively. Mean fetal volume was 979 cm3 (SD 117 cm3) and 2715 cm3 (SD 360 cm3). The best fitting neural network model was attained at 22,000 training iterations with mean DSC 0.925 (SD 0.041). The neural network estimated mean placental volumes at gestational week 27-870 cm3 (SD 202 cm3) (DSC 0.887 (SD 0.034), and to 950 cm3 (SD 316 cm3) at gestational week 37 (DSC 0.896 (SD 0.030)). Mean fetal volumes were 1292 cm3 (SD 191 cm3) and 2712 cm3 (SD 540 cm3), with mean DSC of 0.952 (SD 0.008) and 0.970 (SD 0.040). The time spent for volume estimation was reduced from 60 to 90 min by manual annotation, to less than 10 s by the neural network. CONCLUSION: The correctness of neural network volume estimation is comparable to human performance; the efficiency is substantially improved.