Brief Review — Deep Neural Network for Foreign Object Detection in Chest X-rays

Using **Faster R-CNN** for Foreign Object Detection

Illustrating circle-like foreign objects in CXR (annotating in red). They are, somehow, similar to lung abnormalities, like nodules.

**Deep Neural Network for Foreign Object Detection in Chest X-rays Faster R-CNN for Foreign OO, by The University of South Dakota 2020 CBMS **(Sik-Ho Tsang @ Medium)

Quality Assessment 2018 [5-Layer CNN] 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 ====

• In automated Chest X-Ray (CXR) screening process, foreign objects, such as coins/buttons, medical tubes and devices, and jewelries can adversely impact the performance.
• **Faster R-CNN is used to to detect circle-like foreign objects **of difference sizes in CXRs.

Outline

1. **Faster R-CNN for Foreign Object Detection**
2. Results

1. Faster R-CNN for Foreign Object Detection

1.1. Model Architecture

Faster R-CNN for Foreign Object Detection

• In short, the model has three sections: input, middle, and output.

1. Input data is fed in the input section.
2. The middle section contains three repeated blocks of convolutional, Rectified Linear Unit (ReLU), and pooling layers. The first two convolution layers used 32 filters while 64 for the last one. The filter size, stride, and padding are set to 5×5, 1, and 2, respectively. All the pooling layers have 3×3 filters.
3. The output section consists of five layers: FC layer with 64 neurons, ReLU, FC layer with 10 neurons, Softmax loss layer, and classification layer (using Cross Entropy).

• For optimization, stochastic gradient descent with momentum (SGDM) optimizer, is used.
• Two different setups: a) Without Lung Segmentation (WoLS); and b) With Lung Segmentation (WLS).

1.2. Dataset

• 400 CXRs are from the Indiana CXR dataset maintained by U.S. National Library of Medicine (NLM), National Institutes of Health (NIH).
• In 400 CXRs, there are 1178 circle-like objects, where 325 of them are within the lung region.
• 175 CXRs are used for training and remaining 225 CXRs are used for testing.

2. Results

Ablation of Lung Segmentation (WoLS vs WLS)

Using WLS experimental set up, the proposed tool is better than the best of all (from the literature) by 7% in precision and 2% in F1 score.

SOTA Comparisons (WoLS)

SOTA Comparisons (WLS)

The results outperformed the state-of-the-art techniques.

Illustrative Examples

Some examples are shown as above.