MPS: MRI Pulse Sequence Classification Model

model parameter

https://drive.google.com/drive/folders/176T0DLI5cb2faLDeFzmYNFvcFP0uT2IY?usp=drive_link

Data

Transform

Base

CenterSpatialCropd(keys=["image"], roi_size=(128, 128))

ResizeD(keys=["image"], spatial_size=(224, 224))

NormalizeIntensityD(keys=["image"], nonzero=False, channel_wise=True)

Augment

RandGaussianNoised(keys=["image"], prob=0.45, std=0.09)

RandAdjustContrastD(keys=["image"], gamma=(0.5, 1.5), prob=0.45)

RandGaussianSmoothD(keys=["image"], sigma_x=(0.5, 1.5), prob=0.45)

RandAffined(keys=["image"], prob=0.5, rotate_range=0.1, scale_range=0.1)

RandHistogramShiftd(keys=["image"], num_control_points=10, prob=0.4)

Train

optimizer = optim.Adam( model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay )

criterion = nn.NLLLoss()

scheduler = CosineAnnealingLR(optimizer, len(train_loader) * self.num_epochs)

epochs = 50 / Early Stopping at 23 Epochs

Performance

Loss : 0.0803 / Acc : 0.9779

Accuracy per class

T1 : 100%

T2 : 99.55%

T1 Contrast-Enhenced : 93.39%

FLAIR : 98.57%