Deep Learning Computed Tomography【論文簡單翻譯+理解】
abstract
證明了影象重建能用神經網路表示,具體地說,filtered back-projection(就是CT那個)能夠被對映到深度神經網路結構。迭代重建直接用矩陣相乘效果不好,把back-projection層當作fixed function並把梯度當作投影操作更有效。這樣就可以找到一個數據驅動的方式在投影域和影象域做joint optimization of correction steps。證明了能夠學習weightings和additional filter layers來不斷減少重建誤差(在一個有限的重建角度by a factor of two【自由度2?】),此時計算複雜度和filtered back projection是相同的。
introduction
提出back-projection layer for parallel-beam and fan-beam projection as well as a weighting layer(然後就能端到端訓練)能夠學習back projection之前的project domain的heuristic。因此可以把啟發式的修正方法替換為學習得到的優化策略,還能同時用其他的修正策略。
methodology
離散FBP:
\[f(u, v) \approx \frac{\pi}{N} \sum_{n=1}^{N} q\left(u \cos \left(\theta_{n}\right)+v \sin \left(\theta_{n}\right), \theta_{n}\right) \]1D插值:
設定bias=0且啟用函式是$ f(x)=x\(時,常見的神經元是\)f\left(y_{i}\right)=\sum_{j=1}^{N} w_{i j} x_{j}$,改一下未知數的含義,把f(y)變成f(x,y)代表pixel(x,y),得到:
同樣做插值(the interpolation size big enough to cover the length of the detector by zero-padding the signal as needed):
\[f\left(u_{i}, v_{j}\right) \approx \frac{\pi}{N} \sum_{n=1}^{N} \sum_{m=1}^{M} w_{m}\left(u_{i}, v_{j}, \theta_{n}\right) \cdot q_{m, n} \]線性插值->will yield only up to two non-zero coefficients for every M interpolation coefficients resulting in an extremely sparse matrix.
最終得到的是:
\[f\left(x_{i}, y_{j}\right)=\max \left[0, \sum_{n=1}^{N} \sum_{m=1}^{M} \frac{\pi}{N} w_{m}\left(u_{i}, v_{j}, \theta_{n}\right) \cdot\left(\sum_{k=-M / 2}^{M / 2} w_{k} \cdot p_{m-k, n}\right)\right] \]把FBP的每一步都對應到網路中的一層:
Parallel-Beam Back-Projection Layer(平行光束)
全連線層:This layer has no adjustable parameters.在forward pass中係數被計算出來,\(\mathbf{y}_{l}=\mathbf{W}_{l} \mathbf{y}_{l-1}\),反向傳播時\(\mathbf{E}_{l-1}=\mathbf{W}_{l}^{T} \mathbf{E}_{l}\)。【那麼,訓練的物件其實是“filtered”的部分嗎】
the back-projection operator is the transpose(轉置) of the projection operator.【從公式也許可以看出】
Extension to Fan-Beam Reconstruction(扇形波束)
Weighting Layer 非常稀疏【對角陣?】
Fan-Beam Back-Projection Layer 【一小段話,但是沒看懂】
Convergence and Overfitting
the convolutional layer uses the ramp filter, and the weighting layer accounts for cosine-weighting or redundancy weighting.
These initializations can be used for a very effective pre-training.
experiment & conclusion
沒什麼有意思的。
個人理解:其實是在train濾波的kernel,其他部分基本上都是fixed。不一定對。關於CT的基礎知識可以看上一篇隨筆。對於第二個扇形光束的情況,其實對應weighting layer對應的就是從扇形->平行的過程,convolution layer對應的filter,然後FCL對應backprojection。最後那個線性整形沒理解,是為了去除小於0的值嗎?