In recent years machine learning techniques have seen a tremendous increase in adoption, initially fueled by the massive use of social media leading to very large databases. A development which has also translated to the medical arena. For clinical applications, however, the sizeable data collections machine learning demands remains a challenge. Categorizing machine learning applications in the cathlab, and structurally investigating their respective data needs, aids in developing a systematic approach to the data collection and algorithm development challenges.

Categories

Machine learning applications in the cathlab can be divided in four data categories, dependent on the type of data they receive as input:

1. Image-based: This comprises interventional X-ray, ultrasound, transesophageal echocardiography, intravascular imaging, such as IVUS and OCT, etc.
2. 1D signals: e.g., ECG, respiratory, blood pressure, intravascular measurements such fractional flow reserve FFR, etc.
3. Natural language processing: sources can be either audio fed speech to text (including voice commands and voice annotation), diagnostic patient reports, etc.
4. Hybrid or other data sources: e.g., combination of the input data types above (such as e.g., x-ray images, ECG and respiratory signals).

Machine learning algorithms, regardless of their input data type, typically address a fine-grained task, such as:

• intra-vascular device detection: such as catheter segmentation [1] (Figure 1), needle [2], valve (Figure 2), or stent (Figure 3) detection.
• signal quality improvement: e.g., noise reduction [3].
• image registration: which can be subdivided into rigid and elastic registration [4], and into 2D-3Dand 3D-3D registration.
• event detection: such as adverse event detection [5], or valve deployment (Figure 2).
• procedure phase recognition: which segments the procedure into different time segments [6].
• unstructured data to structured data translation: such as the mining of unstructured text sources [5],
• etc.

These fine-grained algorithmic blocks then feed into high level applications, such as:

• device navigation [1,2],
• lesion quantification [7],
• patient risk stratification [8],
• integrating pre-interventional planning data,
• functional parametrization (e.g. blood flow quantification [9]),
• etc.