Brief Review — Automatic Left Ventricular Outflow Tract Classification for Accurate Cardiac MR Planning

5-Layer CNN for Left Ventricular Outflow Tract (LVOT) Classification

Presence of LVOT Causes Inaccurate Cardiac MR Planning

**Automatic Left Ventricular Outflow Tract Classification for Accurate Cardiac MR Planning 5-Layer CNN, by King’s College London, Hospital NHS Foundation Trust, and Imperial College London 2018 ISBI **(Sik-Ho Tsang @ Medium)

Quality Assessment 2020 [Mean Teacher + ROI Consistency] 2021 [Model Fusion + SVR] 2022 [Swin-MIQA] ==== My Healthcare and Medical Related Paper Readings ==== ==== My Other Paper Readings Are Also Over Here ====

• One result of inaccurate cardiac MR planning is an ‘off-axis’ orientation of the 4-chamber view, often recognized by the presence of the left ventricular outflow tract (LVOT).
• A 5-layer CNN is proposed to automatically detect the presence of the LVOT in cardiac MRI.

Outline

1. 5-Layer CNN
2. Results

1. 5-Layer CNN

• The proposed framework of using a CNN for LVOT detection is as follows

1. Contrast normalization of the target images: Given a full CMR image we first normalize the pixel values between 0 and 1.
2. Image cropping using atrial masks: A template-matching based region of interest (ROI) extraction is used to crop a 64×64 block. 100 MR images are independent from LVOT detection algorithm validation dataset, are solely used for ROI extraction.
3. CNN image classification of correct versus off-axis planning of the 4-chamber images: A CNN of 64×64 — 60×60×32 — 26×26×64–128–2–1 structure, is used.

• The first layer is a convolutional layer, which filters the input with 32 kernels, each of size 5×5. The convolutional layer produces 32 feature maps, each of size 60×60, followed by a 2×2 pooling operation that reduces each feature map to one max.
• Two fully connected (dense) layers (one with 128 and one with 2 nodes) come after the pooling.
• The last layer is a simple softmax classification layer with a one dimensional output that predicts the class.
• **ReLU** is used.
• **Dropout **with a probability of 0.5 at all convolutional layers and after the first fully connected layer to enforce regularization.
• Data augmentation of random rotation and translation is used.

2. Results

Mean accuracy of image classification for 4-chamber LVOT images.

• 82.6% accuracy is obtained by the proposed CNN with data augmentation.