Medical image segmentation basics in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding segmentation masks

In medical image segmentation, what does a segmentation mask represent?

A A binary or multi-class image labeling each pixel with a class such as organ or background

B A 3D model of the organ reconstructed from the scan

C A heatmap showing the probability of disease presence

D A grayscale image showing the intensity of the original scan

💡 Hint

Think about how each pixel is classified in segmentation.

A segmentation mask labels each pixel to indicate which part of the image belongs to which class, such as different organs or tissues.

❓ Predict Output

intermediate

Output shape of segmentation model

Given a 2D medical image input of shape (128, 128, 1), what is the output shape of a segmentation model that predicts 3 classes per pixel?

Computer Vision

input_shape = (128, 128, 1)
num_classes = 3
# Model outputs per-pixel class probabilities

A (3, 128, 128)

B (128, 128, 1)

C (128, 128, 3)

D (128, 3)

💡 Hint

The output shape matches the input height and width, with channels equal to number of classes.

The model outputs a probability for each class at every pixel, so the output shape is (height, width, number_of_classes).

❓ Hyperparameter

advanced

Choosing loss function for imbalanced classes

In medical image segmentation, some organs occupy very small areas compared to background. Which loss function is best to handle this class imbalance?

A Mean Squared Error (MSE)

B Dice Loss

C Categorical Cross-Entropy

D Hinge Loss

💡 Hint

Look for a loss that focuses on overlap between predicted and true masks.

Dice Loss measures overlap and is robust to class imbalance, making it ideal for segmenting small organs.

❓ Metrics

advanced

Evaluating segmentation quality

Which metric best measures how well the predicted segmentation overlaps with the ground truth mask?

A Accuracy

B Recall

C Precision

D Intersection over Union (IoU)

💡 Hint

This metric compares the size of the intersection to the size of the union of two sets.

IoU measures the ratio of the overlap area to the combined area of predicted and true masks, reflecting segmentation quality.

🔧 Debug

expert

Diagnosing model output shape error

You trained a segmentation model but the output shape is (126, 126, 2) instead of (128, 128, 2) for 2 classes. What is the most likely cause?

Computer Vision

model = Sequential([
    Conv2D(32, 1, activation='relu', input_shape=(128,128,1)),
    Conv2D(2, 3, activation='softmax')
])
output = model.predict(input_image)
print(output.shape)

A Last Conv2D layer lacks padding='same', causing shape reduction

B Missing a Flatten layer before the last Conv2D

C Last Conv2D layer uses kernel size 1 but no channels dimension in output

D The model is missing a final activation layer

💡 Hint

Check if convolution reduces spatial dimensions unexpectedly.

Without padding='same', convolution reduces height and width, causing output shape to lose spatial dimensions.