Disease Progression Detection via Deep Sequence Learning of Successive Radiographic Scans

. 2022 Jan 2;19(1):480. doi: 10.3390/ijerph19010480

Algorithm 1: Progression Estimation Function

Input:
F_ref = The reference normal CXR
F_prev = The previous CXR
F_curr, = The current CXR
d = The duration in days between the two successive CXRs
ageRF: Age risk factor computed from historical data using Equation (10)

PEF (F_ref, F_prev, F_curr, d, ageRF)

Extract features from all three input scans to construct feature vectors F_ref, F_prev, and F_curr using

F_{X} = f_{C h e x N e t} (X)

Compute inverse cosine similarity between the feature vectors as

D_{C P} = 1 - cosine_sim (F_{c u r r}, F_{p r e v})

D_{C R} = 1 - cosine_sim (F_{c u r r}, F_{r e f})

D_{P R} = 1 - cosine_sim (F_{p r e v}, F_{r e f})

Compute the relative difference between successive CXRs using the weighted summation

D i f f = \frac{w_{1} . | D_{C R} - D_{P R} | + w_{2} . D_{C P}}{d}

The Diff is then added or subtracted to the previous profile score based on the prediction of the sequence model, or the score remain unchanged if no change is reported.

S_{t} = {\begin{cases} S_{t - 1} - D i f f + \frac{1}{t} \sum_{b = 1}^{t} S_{b} \times a g e R F, f_{S e q} (X_{t - 1}, X_{t}) = “ I m p r o v e ” \\ S_{t - 1} + D i f f + \frac{1}{t} \sum_{b = 1}^{t} S_{b} \times a g e R F, f_{S e q} (X_{t - 1}, X_{t}) = “ D e t e r ” \\ S_{t - 1}, f_{S e q} (X_{t - 1}, X_{t}) = “ N o C h a n g e ” \end{cases}}

Return S_t