Medical Imaging Fundamentals: Image Formation, DICOM, and Quality Metrics

Module: Healthcare AI | Difficulty: Advanced

X-ray Attenuation (Beer-Lambert Law)

where is the incident intensity and is the linear attenuation coefficient at position .

Hounsfield Unit Conversion (CT)

Signal-to-Noise Ratio (SNR)

Contrast-to-Noise Ratio (CNR)

Modulation Transfer Function (MTF)

Common Imaging Modalities

| Modality | Physics | Resolution | Typical Use | |----------|---------|------------|-------------| | X-ray | Photon attenuation | 0.1-0.5 mm | Bone, chest | | CT | X-ray tomography | 0.5-1.0 mm | Trauma, oncology | | MRI | Nuclear magnetic resonance | 0.5-1.5 mm | Brain, soft tissue | | Ultrasound | Acoustic reflection | 0.5-2.0 mm | Obstetrics, cardiac | | PET | Positron emission | 4-5 mm | Oncology, neurology |

import numpy as np
import pydicom
import os

def load_dicom_series(directory):
    slices = []
    for f in os.listdir(directory):
        if f.endswith('.dcm'):
            ds = pydicom.dcmread(os.path.join(directory, f))
            slices.append(ds)
    slices.sort(key=lambda s: s.InstanceNumber)
    volume = np.stack([s.pixel_array for s in slices], axis=-1)
    pixel_spacing = slices[0].PixelSpacing
    slice_thickness = slices[0].SliceThickness
    return volume, pixel_spacing, slice_thickness

def compute_snr(volume, roi_mask):
    signal = np.mean(volume[roi_mask])
    noise = np.std(volume[roi_mask])
    return signal / noise if noise > 0 else float('inf')

def compute_cnr(volume, mask_a, mask_b):
    mu_a, sigma_a = np.mean(volume[mask_a]), np.std(volume[mask_a])
    mu_b, sigma_b = np.mean(volume[mask_b]), np.std(volume[mask_b])
    return abs(mu_a - mu_b) / np.sqrt(sigma_a**2 + sigma_b**2)

def compute_mtf(point_spread_function):
    mtf_1d = np.abs(np.fft.fftshift(np.fft.fft(point_spread_function)))
    mtf_1d /= mtf_1d.max()
    return mtf_1d

print(f'Volume shape: {volume.shape}, Spacing: {spacing}, Thickness: {thickness}')

Research Insight: Deep learning models can learn optimal image representations that surpass handcrafted filters for downstream tasks. Self-supervised pre-training on large unlabeled DICOM datasets enables models to learn clinically meaningful features without explicit annotation, reducing the burden on radiologists during data curation.