The paper introduces a novel framework called Digital Deep Joint Source-Channel Coding (D²-JSCC) for semantic communications (SemCom), which aims to enhance the efficiency and performance of image transmission in 6G applications. D²-JSCC integrates deep learning and digital channel coding to jointly optimize source and channel codings, reducing end-to-end (E2E) distortion. The key contributions include: 1. **D²-JSCC Architecture**: The framework combines deep source coding with an adaptive density model to exploit semantic features and digital channel block coding to protect encoded features from channel distortion. The E2E distortion is characterized using a Bayesian approximation and a Lipschitz assumption on deep neural networks (DNNs). 2. **E2E Distortion Approximation**: The E2E distortion is approximated as a function of DNN parameters and channel rate, facilitating the optimization process. 3. **Optimal Rate Control**: An efficient two-stage algorithm is proposed to balance the trade-off between source and channel rates. In the first step, a look-up table of DNN models is created to select the optimal source encoder based on the given channel SNR. In the second step, the selected DNN model is retrained to adapt to the channel SNR, optimizing the E2E performance. 4. **Experimental Results**: Experiments demonstrate that D²-JSCC outperforms classic deep JSCC, mitigating the "cliff effect" and "leveling-off effect" commonly seen in digital systems. The proposed framework also shows improved performance with increasing block length, approaching the capacity-achieving code performance. The paper addresses the challenges of optimizing DNN parameters and channel coding in a discrete system, providing a practical solution for efficient and reliable image transmission in SemCom.The paper introduces a novel framework called Digital Deep Joint Source-Channel Coding (D²-JSCC) for semantic communications (SemCom), which aims to enhance the efficiency and performance of image transmission in 6G applications. D²-JSCC integrates deep learning and digital channel coding to jointly optimize source and channel codings, reducing end-to-end (E2E) distortion. The key contributions include: 1. **D²-JSCC Architecture**: The framework combines deep source coding with an adaptive density model to exploit semantic features and digital channel block coding to protect encoded features from channel distortion. The E2E distortion is characterized using a Bayesian approximation and a Lipschitz assumption on deep neural networks (DNNs). 2. **E2E Distortion Approximation**: The E2E distortion is approximated as a function of DNN parameters and channel rate, facilitating the optimization process. 3. **Optimal Rate Control**: An efficient two-stage algorithm is proposed to balance the trade-off between source and channel rates. In the first step, a look-up table of DNN models is created to select the optimal source encoder based on the given channel SNR. In the second step, the selected DNN model is retrained to adapt to the channel SNR, optimizing the E2E performance. 4. **Experimental Results**: Experiments demonstrate that D²-JSCC outperforms classic deep JSCC, mitigating the "cliff effect" and "leveling-off effect" commonly seen in digital systems. The proposed framework also shows improved performance with increasing block length, approaching the capacity-achieving code performance. The paper addresses the challenges of optimizing DNN parameters and channel coding in a discrete system, providing a practical solution for efficient and reliable image transmission in SemCom.